Free tool · About 12 minutes
AI Search Visibility Audit
Score your website on 50 checks that decide if ChatGPT, Perplexity, Gemini and Claude can find, read and quote you. Mark each check. Your score changes as you go. Then get a ranked list of the fixes that earn the most points for the least work.
- 50 checks in four phases, ordered by the work each one takes
- Phases 1 and 2 hold 27 checks and 54 of the 100 points
- No account needed. Your answers stay in your browser
How to use the audit
- Mark each check: No, Partly or Yes. Each check is worth 2 points.
- Open "How to do this" on any check you are not sure about. It shows why the check matters, five steps, a test for done and the common mistake.
- Read "Do these next". It ranks your open checks by points per day of work.
- Get your scorecard. It is a one-page report you can download or save as a PDF.
Phase 1 of 4
Quick wins
Settings, files and profiles. You change nothing about your content. One person can clear this list in a working day.
-
01
Unblock AI crawlers in robots.txt
Check that robots.txt does not block GPTBot, ClaudeBot, PerplexityBot or CCBot.
2 points · Estimate: 15 min
How to do this
Why it matters
One Disallow line can hide the whole site from AI answer engines. Most of these blocks are accidents. They arrive with a security plugin or a CDN setting that nobody remembers approving.
How to do it
- Open yoursite.com/robots.txt in a browser and read it.
- Look for rules that name GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-SearchBot, PerplexityBot, Perplexity-User, Google-Extended, Applebot-Extended or CCBot.
- Decide per crawler. Search crawlers fetch a page to answer a live question. Training crawlers collect data to train a model. You can allow the first group and block the second.
- Check the firewall and the CDN as well. Cloudflare, Wordfence and similar tools block AI bots at the edge, so robots.txt can look clean while the request never arrives.
- Recheck after any plugin update. Security plugins reset these rules.
You are done when
Your server logs show 200 responses to the crawlers you chose to allow.
Common mistake
Blocking every AI user agent to protect the content, then asking why the brand never appears in answers.
-
02
Submit your sitemap to Bing Webmaster Tools
Bing feeds several answer engines. Add the site and submit the sitemap.
2 points · Estimate: 30 min
How to do this
Why it matters
ChatGPT search draws on the Bing index. A page Bing has not crawled cannot be retrieved to answer a question, whatever it ranks in Google.
How to do it
- Create a free Bing Webmaster Tools account. Import the site from Google Search Console to save the setup.
- Verify the domain, then submit yoursite.com/sitemap.xml.
- Run the URL Inspection tool on your five most important pages and confirm each one is indexed.
- Turn on IndexNow so new and changed pages get pushed the moment you publish.
- Check the crawl errors report once a month.
You are done when
A site:yourdomain.com search in Bing returns your key pages.
Common mistake
Assuming Google coverage means Bing coverage. The two indexes differ, sometimes by a lot.
-
03
Publish llms.txt at your site root
One plain text file that points AI crawlers at your most important pages.
2 points · Estimate: 2 hours
How to do this
Why it matters
llms.txt is a markdown file at /llms.txt. It tells an AI system which pages matter and what each one covers, in one place, without crawling the whole site.
How to do it
- Write it in markdown. Start with an H1 of the brand name.
- Add a blockquote of one sentence that says what you do.
- Add H2 sections such as Products, Documentation, Pricing, About.
- Under each section, list links as - [Page name](full URL): one line on what it covers.
- Upload it to the site root so it resolves at yoursite.com/llms.txt. Keep it short and link only the pages you want quoted.
You are done when
The file loads as plain text in a browser at the root URL.
Common mistake
Publishing a link dump with no descriptions. The one line per link is the part that does the work.
-
04
Add basic schema to your key pages
Organization, FAQPage and Article, on the pages that matter most.
2 points · Estimate: 4 hours
How to do this
Why it matters
Structured data states the facts instead of leaving them to be inferred. It will not force a citation on its own, but without it a system has to guess what the page is and who published it.
How to do it
- Add Organization schema once, on the home page. Include the legal name, logo, URL and sameAs links to your profiles.
- Add Article or BlogPosting to posts, with headline, author, datePublished and dateModified.
- Add FAQPage only where a real FAQ block is visible on the page.
- Add Product, Service or LocalBusiness where they apply.
- Test every page in the Rich Results Test and the Schema.org validator.
You are done when
Both validators return no errors on your top ten pages.
Common mistake
Marking up an FAQ a visitor cannot see. The visible content and the structured data have to match.
-
05
Show a visible last updated date
Put the date on the page. Change it when you change the content.
2 points · Estimate: 2 hours
How to do this
Why it matters
Answer engines favor recent content. A page with no date reads as undated, and an undated page loses to a dated one on the same subject.
How to do it
- Put the date in the page body, near the title, in plain text a reader can see.
- Set dateModified in the Article schema to the same date.
- Change the date only when the content changes. Touching a file does not count.
- Start with your top twenty pages, then extend the habit.
- Record what changed in the review log. It makes the next review faster.
You are done when
A reader can see when the page last changed without opening the source.
Common mistake
Auto-stamping today's date on every page. The signal stops meaning anything, and readers notice.
-
06
Make your name and description identical everywhere
Same business name, same address, same one line description on every profile.
2 points · Estimate: 4 hours
How to do this
Why it matters
AI systems build one profile of your brand out of many sources. Conflicting names, addresses and descriptions split that profile into several weaker ones.
How to do it
- Write one canonical block: legal name, trading name, one line description, address, phone, founding year, URL.
- Fix your own site first. Footer, about page, contact page, Organization schema.
- Work through LinkedIn, Crunchbase, G2, directories, partner pages and old press kits.
- Keep the block in a shared document and send it to anyone who writes about you.
- Re-audit twice a year. Old profiles drift.
You are done when
Ten searches for your brand return the same description everywhere.
Common mistake
Rewriting the boilerplate for each channel. Variety helps humans and hurts entity resolution.
-
07
Get listed in the directories your industry trusts
G2, Capterra, Clutch, trade association lists. Claim the profile and complete it.
2 points · Estimate: 4 hours
How to do this
Why it matters
Most brand mentions inside AI answers come from third-party pages, not from your own site. Directories are the cheapest third-party pages you can earn.
How to do it
- List the directories your buyers actually check, plus the member list of any association you belong to.
- Claim the profile rather than leaving the auto-generated stub.
- Complete every field: category, description, pricing, integrations, logo, screenshots.
- Use the canonical description from the previous check, word for word.
- Aim for breadth of source types, not more pages on one site.
You are done when
Your brand appears on at least five independent source types.
Common mistake
Claiming a profile and leaving it half filled. An empty profile still ranks and still describes you badly.
-
08
Claim your Wikidata and Google Business Profile entries
Correct the facts AI systems already hold about your brand.
2 points · Estimate: 2 hours
How to do this
Why it matters
These are reference sources that AI systems read directly. Wrong facts there follow you into every answer about the company.
How to do it
- Search Wikidata for your brand. Claim the item, or create it if none exists.
- Add founding date, industry, official website, headquarters and official identifiers.
- Claim the Google Business Profile. Correct the category, hours, address and description.
- Do not write your own Wikipedia article. It needs independent coverage and it will be deleted. Earn the coverage first.
- Recheck each quarter. Anyone can edit these records.
You are done when
Your knowledge panel shows the facts you would state yourself.
Common mistake
Editing Wikipedia as the company. Conflict of interest edits get reverted and the reversal is public.
-
09
Track how often you appear in AI answers
Ask the same ten buyer questions each month. Record when your brand appears.
2 points · Estimate: 1 hour
How to do this
Why it matters
You cannot improve a number nobody records. Most teams have never looked, so the first month of data is usually the most useful.
How to do it
- Write ten to twenty buyer questions in the words a buyer uses, not your product name.
- Run them monthly in ChatGPT, Perplexity, Gemini and Claude, in a fresh session with memory and personalization off.
- Record four things per question: were you named, were you linked, who else was named, which pages were cited.
- Keep it in one sheet with a row per question per month.
- Read the competitor column first. It tells you which pages to go and get on.
You are done when
You can say what share of your buyer questions named your brand this month.
Common mistake
Asking what your brand is. That tests recognition, not visibility. Buyers do not search your name.
-
10
Give reporters a quote and a number on request
Keep one approved quote and one sourced statistic ready to send.
2 points · Estimate: 2 hours
How to do this
Why it matters
Third-party coverage is what answer engines borrow authority from. Reporters need a usable quote fast, and speed usually decides who gets included.
How to do it
- Build a one page kit: two approved quotes with name and title, two sourced statistics, one boilerplate paragraph, headshots and logo.
- Answer inside the reporter's deadline, not inside your review cycle.
- Give one number with a date and a method. A number nobody can source gets cut.
- Watch the journalist request services and the request hashtags on LinkedIn.
- Log every placement. It becomes the target list for Phase 3.
You are done when
You can answer a media request in under an hour without calling a meeting.
Common mistake
Sending marketing copy. A reporter needs a sentence they can print with your name after it.
-
11
Check that no cookie wall or login hides your content
A consent banner, paywall or sign-in gate can hide the page from a crawler.
2 points · Estimate: 1 hour
How to do this
Why it matters
A crawler that meets a consent wall, a hard paywall or a login screen sees the wall, not the page. Your best content can be invisible while the page looks correct to a person.
How to do it
- Open your key pages in a private window with no cookies set. Read what appears before you accept anything.
- Fetch the same pages with a plain request and read the raw HTML. Compare it with what a visitor sees.
- Set consent so the page text loads before a choice is made. Gate the tracking, not the words.
- Move content you want quoted in front of the form. Gate the deeper asset instead.
- Check the mobile app prompt as well. Some interstitials cover the whole page.
You are done when
A request with no cookies and no session returns the full page text.
Common mistake
Gating the article and the download together. Gate the download, publish the argument.
-
12
Fix crawl errors and broken links
A crawler that meets errors reads less of your site.
2 points · Estimate: 2 hours
How to do this
Why it matters
Errors, dead links and redirect loops spend the crawl on nothing. The pages behind them get read late, or not at all.
How to do it
- Read the crawl error reports in Google Search Console and in Bing Webmaster Tools.
- Fix or redirect every 404 that still has links pointing at it.
- Remove redirect chains. One hop from the old URL to the final page.
- Confirm your important pages return 200, not a 3xx or a 5xx.
- Crawl the site once a quarter and clear the new errors.
You are done when
Both webmaster tools report no errors on your key pages.
Common mistake
Redirecting every dead URL to the home page. That tells a crawler the page is gone and gives the reader nothing.
-
13
Correct the wrong facts AI answers give about you
Trace each wrong claim to its source page and fix it there.
2 points · Estimate: 2 hours
How to do this
Why it matters
An answer engine repeats what its sources say. A wrong price, a dead product name or an old founder story stays wrong until you fix the page it came from.
How to do it
- In your monthly tracking, write down every claim about you that is wrong or out of date.
- Ask the engine which sources it used, then open them.
- Fix the ones you control first: your own pages, your profiles, your directory entries.
- Ask the owner of a third-party page to correct it. Send the correct line, ready to paste.
- Recheck the question next month. A correction takes time to reach an index.
You are done when
Every wrong claim you found last month has a traced source and a correction request.
Common mistake
Complaining about the answer instead of fixing the source. The engine repeats the page, not your opinion of it.
Phase 2 of 4
Content work
Edits to pages you already own, plus the off-site work that earns citations. No rebuild, no migration.
-
14
Add real statistics with sources to existing pages
Replace vague claims with a number, a date and a source.
2 points · Estimate: 2 weeks
How to do this
Why it matters
Adding sourced numbers is one of the strongest single changes tested for AI citation. A specific, sourced claim is a unit an answer engine can lift and reuse whole.
How to do it
- Read your top twenty pages and mark every vague claim: many customers, significant savings, industry leading.
- Replace each one with a number, a date and a source. Your own data counts if you state how you measured it.
- Put the number in the sentence, not only in a chart. A chart image is not read as text.
- Add the year. An undated statistic ages badly and gets skipped.
- Check that the study you cite actually covers your population and your claim before you use it.
You are done when
Each key page carries at least three sourced numbers in the body text.
Common mistake
Citing a study because the words match your point. Check the population and the scope first, or leave the number out.
-
15
Put named expert quotes into your content
A real person with a real title carries more weight than an unsigned line.
2 points · Estimate: 1 week
How to do this
Why it matters
An attributed quote is a clean, self-contained fact with a person attached. Answer engines lift it whole. An unsigned assertion has nothing to attach the credibility to.
How to do it
- Interview your own experts for fifteen minutes and pull two or three quotes per article.
- Use the full name, the job title and the company in the attribution.
- Mark the quote up as a blockquote so its boundaries are clear.
- Quote outside experts where you can get them. Independent voices carry further than your own.
- Keep a quote bank so writers stop inventing generic lines.
You are done when
Every important page carries at least one named, titled quote.
Common mistake
Quoting "our team" or "a customer". Anonymous attribution adds nothing at all.
-
16
Write headings as full questions
Match the way buyers ask, not the way keyword tools write.
2 points · Estimate: 1 week
How to do this
Why it matters
People ask AI systems full questions. Headings shaped like the question help the retrieval step find the right passage inside a long page.
How to do it
- Turn keyword headings into questions. "Pricing" becomes "How much does it cost?"
- Answer the question in the first sentence after the heading. Do not build up to it.
- Use the words buyers use. Pull them from sales calls and support tickets.
- Keep one question per heading. Two questions split the answer in half.
- Leave navigation labels and section titles alone. Use questions where a real question exists.
You are done when
Your H2s read as questions a person would type or say out loud.
Common mistake
Turning every heading into a question, including the ones that are not questions. It reads as a template.
-
17
Add an FAQ block to your most important pages
Answer the question in the first two sentences. Support it below.
2 points · Estimate: 1 week
How to do this
Why it matters
FAQ blocks produce short, self-contained answers, which is the format retrieval prefers. They also cover the questions your main copy skips.
How to do it
- Take the ten to fifteen questions your buyers ask most. Sales calls and support tickets hold the list.
- Answer each one completely in the first two or three sentences, then add supporting detail below.
- Keep each answer under about 100 words so it does not get truncated in the wrong place.
- Add FAQPage schema that matches the visible questions exactly.
- Review the list twice a year. The questions change as the product changes.
You are done when
A reader gets a usable answer to each question without scrolling further.
Common mistake
Writing questions nobody asks so you can repeat keywords. Buyers ask about price, setup time, integrations and what happens when it fails.
-
18
Add author bios with real credentials
Name the author. Show the background that makes the answer credible.
2 points · Estimate: 3 days
How to do this
Why it matters
A named person with a background you can verify is worth more in a trust calculation than an "Admin" byline.
How to do it
- Give every author a bio page with a real name, job title, credentials and years of experience.
- Add Person schema and link the bio to LinkedIn and professional profiles with sameAs.
- Put a short bio at the end of each article and link to the full page.
- Remove ghost bylines. If the byline is not a real person, publish under the company instead.
- Say why this person can answer this question. That is the part that counts.
You are done when
Every article names a person a reader can look up.
Common mistake
A bio that lists hobbies. List the reason this person is qualified to answer.
-
19
Earn real Reddit conversations about your brand
Answer questions where your buyers already ask them. Do not astroturf.
2 points · Estimate: ongoing
How to do this
Why it matters
Reddit appears constantly as a source in AI answers, and question threads carry most of it. It works because the threads read as independent opinion.
How to do it
- Find the subreddits where your buyers already ask about your category.
- Have real team members answer with their real accounts, and disclose the affiliation. Most subreddits require it.
- Answer the question first. Mention the product only when it is the actual answer.
- Ask happy customers to share their own experience in their own words. Do not script them.
- Read the rules of each subreddit before the first post.
You are done when
Searching Reddit for your category returns threads where someone other than you discusses your brand.
Common mistake
Astroturfing. Reddit finds it, the thread gets removed, the account gets banned, and the damage outlasts the post.
-
20
Put your experts on Quora and industry forums
Founders and specialists answer real questions under their own names.
2 points · Estimate: ongoing
How to do this
Why it matters
Consulting-led fields see high AI referral traffic, and forum answers are part of the reason. A signed expert answer is easy to quote and easy to trust.
How to do it
- Pick two or three forums where your buyers ask questions, then set up real profiles properly.
- Answer under real names with real titles.
- Write complete answers. A three line answer with a link gets ignored and often removed.
- Budget one hour a week per expert. Consistency beats volume.
- Answer old questions too. They keep getting read.
You are done when
A search for your category returns forum answers signed by your people.
Common mistake
Dropping links. Answer the question so well that the link is optional.
-
21
Build review volume on G2, Trustpilot and category sites
Ask happy customers. Volume and recency both count.
2 points · Estimate: ongoing
How to do this
Why it matters
Answer engines treat review volume and sentiment as a trust signal. Reviews are third-party statements about you, which is the source type these systems prefer.
How to do it
- Pick the two or three platforms your buyers actually read. Depth on two beats presence on eight.
- Ask at a natural moment: after onboarding, after a support win, after a renewal.
- Never pay for a positive review. Any incentive must be unconditional and disclosed.
- Reply to every review, including the critical ones. The reply gets read as well.
- Keep a steady monthly rhythm rather than one campaign a year.
You are done when
New reviews arrive every month rather than in one burst.
Common mistake
One review push a year. Recency counts, and a wall of reviews from the same week reads as a campaign.
-
22
Pitch for best of and comparison roundups
Third party lists are the pages AI answers quote most.
2 points · Estimate: 2 weeks
How to do this
Why it matters
Roundups and comparison pages are what answer engines quote when a buyer asks who the options are. You cannot appear in a list you are not on.
How to do it
- Run your buyer questions and record which roundup pages get cited. Those are your targets.
- Pitch the writer with what makes you different, plus one number and one customer example.
- Offer a free account so the writer can actually test the product.
- Ask to be added to existing lists, not only to new ones. Updates are easier to win.
- Recheck twice a year. Lists get refreshed and brands drop off quietly.
You are done when
Your brand appears on the roundup pages that AI answers already cite.
Common mistake
Pitching your press release. The writer needs a reason you belong in the comparison.
-
23
Distribute through LinkedIn, YouTube and podcasts
Transcripts get read. One post becomes four sourceable formats.
2 points · Estimate: ongoing
How to do this
Why it matters
Video transcripts and podcast text are read as citable sources. One piece of work you already published becomes four sources with no new writing.
How to do it
- Take a published article and record a short video walkthrough. Publish it with a full transcript.
- Post the core argument as a LinkedIn article, not only as a status update.
- Pitch the same argument as a podcast topic and ask the host to publish a transcript.
- Keep the wording of your key claims the same across formats. Repetition helps entity resolution.
- Link every format back to the page that owns the topic.
You are done when
Your main argument exists in four places under your name.
Common mistake
Publishing video with no transcript. Untranscribed video is not readable and cannot be quoted.
-
24
Publish comparison and alternative pages
Buyers ask AI who the options are. Answer that on your own site.
2 points · Estimate: 2 weeks
How to do this
Why it matters
Comparison questions are where buyers decide. If you have no page that compares you honestly, the only comparison an answer engine can quote is written by somebody else.
How to do it
- List the competitors your buyers name on sales calls. Build one page for each pair.
- Give every page a comparison table that uses the same fields for both sides.
- Say plainly who each option suits best, including the cases where you are the wrong choice.
- Add an alternatives page for the category leader. That question has real volume.
- Date the facts and keep them current. A wrong competitor price discredits the whole page.
You are done when
Every competitor a buyer names has a page on your site that compares you honestly.
Common mistake
Writing a comparison where you win every row. Readers and answer engines both discount it.
-
25
Publish your prices
"What does it cost" is a top buyer question. Answer it in public text.
2 points · Estimate: 3 days
How to do this
Why it matters
A price behind a form cannot be quoted. When a buyer asks what something costs, the engines answer with the vendors who published a number.
How to do it
- Publish a price, a range, or a worked example. A range beats silence.
- Put the number in text on the page, not inside an image or a script.
- State what moves the price: seats, volume, term, scope.
- Add Offer or Service schema carrying the same figures.
- Date the page and review it whenever pricing changes.
You are done when
A reader can learn what you cost without talking to anyone.
Common mistake
"Contact us for pricing" on the pricing page. That page answers nothing and gets skipped.
-
26
Format your answers for extraction
Tables, numbered steps and short definitions get lifted whole.
2 points · Estimate: 1 week
How to do this
Why it matters
A clean table or a numbered procedure is a complete answer in one block. A wall of prose has to be cut somewhere, and the cut often ruins the meaning.
How to do it
- Turn comparisons into tables with a header row and one idea per cell.
- Turn procedures into numbered steps, one action per step.
- Give each key term a one sentence definition in its own paragraph.
- Add HowTo schema to real step-by-step content.
- Keep tables in HTML. A table saved as an image cannot be read.
You are done when
Your key pages each carry at least one table or one numbered procedure.
Common mistake
Pasting a screenshot of a spreadsheet. Nothing can read the numbers in a picture.
-
27
Write alt text and publish transcripts
Images and video carry no meaning until you write it down.
2 points · Estimate: 1 week
How to do this
Why it matters
Answer engines cite images and pull from video, and both need words. An unlabeled image and an untranscribed video hold information that nothing can read.
How to do it
- Write alt text that says what the image shows, not what the file is called.
- Give every chart and diagram a caption that states the finding in words.
- Rename image files with descriptive words before you upload them.
- Publish a full transcript with every video and every podcast episode.
- Add ImageObject or VideoObject schema where the media carries the point.
You are done when
Every image on your key pages has alt text, and every video has a transcript.
Common mistake
Alt text that repeats the file name or stuffs keywords. Describe the picture.
Phase 3 of 4
Structural work
New systems, not new pages. You change how the site is organized, linked, described and measured.
-
28
Rewrite key pages as answer first passages
Give the direct answer in the first 40 to 60 words. Support it below.
2 points · Estimate: 4 weeks
How to do this
Why it matters
Retrieval works on passages, not whole pages. A passage that answers the question by itself can be lifted. A passage that depends on the paragraph above it cannot.
How to do it
- Put the direct answer in the first 40 to 60 words after the heading. No preamble.
- Keep each passage self-contained, around 130 to 170 words.
- Repeat the subject instead of writing "it" or "this". A pronoun with no antecedent breaks the passage.
- One idea per passage. Break long sections into headed passages.
- Cut the throat clearing. "In today's fast moving landscape" is not an answer.
You are done when
Any passage on the page still makes sense when it is read alone.
Common mistake
Rewriting the introduction and leaving the rest of the page as it was.
-
29
Build a full entity register, not a few tags
Declare every product, person, location and claim in structured data.
2 points · Estimate: 4 weeks
How to do this
Why it matters
A few schema tags describe pages. An entity register describes the business. It lets a system verify a fact instead of inferring it from wording.
How to do it
- List every entity: products, services, people, locations, events, awards, integrations, named customers.
- Give each entity one canonical page and one stable @id URL.
- Connect them with schema properties. Link a Product to its Organization, a Person to their articles, a Service to its area served.
- Add sameAs on every entity, pointing at the external profiles that confirm it.
- Generate the markup from one source of truth so it cannot drift page by page.
You are done when
Every entity you mention has one page, one identifier and links out to independent confirmation.
Common mistake
Adding schema to pages without connecting them. Unlinked entities describe a page, not a business.
-
30
Merge or remove thin and duplicate pages
Fewer, stronger pages give a clearer signal about what you know.
2 points · Estimate: 2 weeks
How to do this
Why it matters
Forty posts saying similar things give a diluted signal. A system cannot tell which page holds your position, so it may trust none of them.
How to do it
- Export every URL and group them by the question each page answers.
- In each group, keep the strongest page and merge the useful parts of the others into it.
- Redirect every removed URL to the kept page with a 301.
- Clear the thin pages that answer nothing: empty tag archives, stub category pages, dead event pages.
- Resubmit the sitemap once the redirects are live.
You are done when
Each buyer question has exactly one page on your site that owns it.
Common mistake
Deleting pages without redirects. You lose the links that were pointing at them.
-
31
Reorganize the site into topic clusters
Pillar pages with real internal links, not a flat list of posts.
2 points · Estimate: 6 weeks
How to do this
Why it matters
A flat list of posts gives nothing to read structurally. Clusters show which topics you own and how deep the coverage runs.
How to do it
- Pick three to six topics you want to be known for. No more than six.
- Build one pillar page per topic that covers it completely and links to every supporting page.
- Link every supporting page back to its pillar with anchor text that names the topic.
- Rework the navigation to match. A menu that contradicts the clusters cancels them.
- Assign new content to a cluster before it is written.
You are done when
Every page belongs to exactly one cluster and links to its pillar.
Common mistake
Building pillar pages and leaving the internal links alone. The links are the structure.
-
32
Publish llms-full.txt and a markdown mirror
A text first copy of your key content that AI systems can read end to end.
2 points · Estimate: 2 weeks
How to do this
Why it matters
llms.txt is an index. llms-full.txt is the content itself, in one clean text file, with no navigation, scripts or styling in the way.
How to do it
- Generate llms-full.txt from your key content, in markdown, in a sensible order.
- Serve a markdown version of each page as well. A common pattern is /page.md beside /page.
- Automate the build so the mirror regenerates on publish.
- Keep the mirror identical to the live page. A mirror that disagrees with the page is a contradiction, not a help.
- Link the mirror from llms.txt so it can be found.
You are done when
The mirror regenerates on publish with no manual step.
Common mistake
Treating it as a one-off export. A hand-maintained copy goes stale in a month, and a stale copy is worse than none.
-
33
Commission original research
Produce a statistic that other sites have to credit back to you.
2 points · Estimate: 8 weeks
How to do this
Why it matters
A number only you have is the number other sites cite. Those citations are third-party pages pointing at you, which is the strongest form of the signal.
How to do it
- Pick a question your industry argues about and has no data for.
- Choose a method you can defend: a survey with a stated sample, an analysis of your own aggregated data, or a repeated benchmark.
- Publish the method, the sample size, the dates and the limits. A study with a hidden method gets ignored.
- Give the finding a plain name people can quote in one line.
- Republish it annually. The second year turns a study into a benchmark.
You are done when
Other sites cite your figure by name.
Common mistake
Surveying fifty of your own customers and calling it industry research. State exactly who you asked.
-
34
Run a real digital PR program
A standing program with targets and a calendar, not occasional pitching.
2 points · Estimate: ongoing
How to do this
Why it matters
Consistent third-party coverage is a program with a budget line, not a favor from a contact. Occasional pitching gives occasional results.
How to do it
- Set a target: a number of placements per quarter on a named list of publications.
- Build a story calendar: data stories, expert commentary, customer results, category arguments.
- Give one person the target and the budget. Shared ownership means no ownership.
- Track which placements show up in AI answers, not only which ones send traffic.
- Feed the media kit from Phase 1 into every pitch.
You are done when
Coverage arrives on a schedule you set rather than by chance.
Common mistake
Measuring PR by traffic alone. Many of the most valuable citations send almost no clicks.
-
35
Rebuild internal links around entities
Link because two pages describe the same product or person, not because of a keyword.
2 points · Estimate: 3 weeks
How to do this
Why it matters
Keyword links say what a page targets. Entity links say how two things relate. The relationship is the part a crawler reads.
How to do it
- Map which pages describe which people, products, concepts and locations.
- Link them because of the relationship. Case study to product, product to integration, integration to partner.
- Use anchor text that names the entity. "Our platform" tells nobody anything.
- Fix orphan pages. A page with no inbound internal link reads as unimportant.
- Check the depth. Anything more than three clicks from the home page needs a better route.
You are done when
You can trace a path from any entity page to every related entity page.
Common mistake
Adding a related posts widget and calling it internal linking. An automated block does not express a relationship.
-
36
Fix analytics so AI referrals are visible
Most AI traffic arrives with no referrer. Tag and segment it or you will not see it.
2 points · Estimate: 1 week
How to do this
Why it matters
A large share of AI-driven traffic arrives with no referrer header and lands in Direct in a default setup. If you cannot see it, you cannot argue for the work.
How to do it
- Build a GA4 exploration that filters session source for chatgpt.com, perplexity.ai, gemini.google.com, claude.ai, copilot.microsoft.com and their variants.
- Add a regex channel group so the traffic reports as one channel over time.
- Read the server logs as well. Logs show crawler visits that analytics never records.
- Segment AI referrals rather than merging them into organic. They behave differently.
- Set a monthly report so the trend is visible, not just the total.
You are done when
You can report AI referral sessions and conversions as their own channel.
Common mistake
Adding a dashboard widget and stopping there. The tagging is the actual work.
-
37
Put a named owner on a freshness calendar
Assign 30, 90 and 180 day review cycles to a person, not to a team.
2 points · Estimate: 3 days
How to do this
Why it matters
Freshness is a signal you have to keep paying. A one-time cleanup decays, and the decay is invisible until the citations stop.
How to do it
- Sort content into three cycles. 30 days for pricing, product and anything with numbers. 90 days for guides and comparisons. 180 days for durable explainers.
- Put one named person on each cycle.
- Define a review: check the facts, refresh the numbers, update the visible date. A cosmetic edit does not count.
- Log the review date and the reviewer.
- Report the overdue count each month. That number is the health of the system.
You are done when
The calendar exists, it has names on it, and last month's reviews are recorded.
Common mistake
Assigning the whole calendar to "marketing". Work with no name attached does not happen.
-
38
Fix crawl hygiene across the whole site
Speed, status codes, canonicals and redirect chains, at scale.
2 points · Estimate: 3 weeks
How to do this
Why it matters
Retrieval favors pages that load fast and resolve cleanly. Slow pages, duplicate URLs and long redirect chains all cut how much of your site gets read.
How to do it
- Meet the Core Web Vitals thresholds on your main templates, not only on the home page.
- Give every page one canonical URL. Resolve parameter, trailing slash and letter case variants.
- Remove redirect chains so every old URL reaches its final page in one hop.
- Cut the weight that delays first render: unused scripts, oversized images, blocking fonts.
- Watch server response time under crawler load, not only under human traffic.
You are done when
Your main templates pass Core Web Vitals and every URL resolves in one hop.
Common mistake
Tuning the home page alone. Crawlers read your templates, and most of your pages are templates.
-
39
Open your documentation and support answers to the public
Move the answers out of the help desk and out of the login.
2 points · Estimate: 6 weeks
How to do this
Why it matters
Your support team already writes the best answers to real questions. Behind a login, none of it can be read or quoted.
How to do it
- Publish product documentation on a public URL, not inside the application.
- Turn your most common support tickets into public help articles, in the buyer's words.
- Run a public community or forum where questions and answers stay readable.
- Link the documentation from llms.txt and from the marketing pages it supports.
- Give the documentation the same freshness cycle as the rest of the site.
You are done when
A buyer can read your documentation and your common answers without an account.
Common mistake
Publishing documentation that only makes sense to existing customers. Write the first paragraph for someone who has not bought.
Phase 4 of 4
Rebuild
Architecture and org design. These change how the organization works, so they need a budget and a sponsor.
-
40
Move to a headless, API driven content stack
Content as data, delivered to any surface, with schema as the identity layer.
2 points · Estimate: 6 months
How to do this
Why it matters
A schema plugin bolted onto a page-based site can only describe pages. A headless stack holds content as structured data and delivers it to any surface, including AI systems.
How to do it
- Separate content from presentation. Content lives in a structured store with typed fields.
- Deliver through an API so the same content serves the site, the app, partners and AI systems.
- Generate schema from the content model rather than writing it per page.
- Make llms.txt, the markdown mirror and the sitemap automated build outputs.
- Model the content before you pick the platform.
You are done when
A change to one field updates the page, the schema, the mirror and the API at once.
Common mistake
Replatforming without a content model. Headless with unstructured fields gives you the same problem on newer infrastructure.
-
41
Rebuild information architecture around questions
Organize the site by what buyers ask, not by what you sell.
2 points · Estimate: 4 months
How to do this
Why it matters
Most sites are organized by what the company sells. Buyers arrive with a question, and retrieval matches the question, not your product taxonomy.
How to do it
- Collect the questions from sales calls, support tickets, search queries and AI prompts. Group them.
- Build the navigation from the groups. Product pages become supporting evidence, not the top level.
- Give every significant question one page that owns it.
- Redirect the old structure properly. This is a migration, not a menu change.
- Expect resistance. Product marketing, sales and support all have a stake in the old structure.
You are done when
A buyer can find the answer to their question without knowing your product names.
Common mistake
Adding a question-led section beside the product-led site. Two structures compete and both get weaker.
-
42
Build a knowledge graph of your business
Every entity you touch, cross referenced and open to verification.
2 points · Estimate: 6 months
How to do this
Why it matters
A graph holds the relationships, not just the records. It can answer a question you never wrote a page for, because the connection between two things is stated.
How to do it
- Model the entities and the relationships between them before you build anything.
- Store the graph as the source of truth. Pages and schema become views of it.
- Give every entity a stable identifier that survives a page move.
- Connect the graph to external identifiers: Wikidata, industry registries, standard codes.
- Put one team on maintenance. A graph nobody updates becomes wrong quickly.
You are done when
You can query the graph for a relationship that no single page states.
Common mistake
Calling a spreadsheet of schema tags a knowledge graph. Without the relationships, it is a list.
-
43
Expose a live query interface such as MCP
Let AI agents pull current data from you instead of a stale copy.
2 points · Estimate: 3 months
How to do this
Why it matters
A static page describes your business as it was when the page was written. A live interface answers with the current price, the current stock, the current availability. It also lets an agent do something rather than read about it.
How to do it
- Decide what an agent should be able to ask: availability, pricing, order status, specifications.
- Build a server that exposes those as tools, with clear names and descriptions.
- Apply authentication, rate limits and an audit log from the first day. This is a public door onto live data.
- Version the interface so changes do not break the agents already using it.
- Publish how to connect. An interface nobody can find gets no use.
You are done when
An agent can answer a current-state question about your business without a human.
Common mistake
Exposing the whole database. Expose the questions you want answered, and nothing else.
-
44
Serve clean HTML to crawlers first
Cut the client side rendering that hides your content from crawlers.
2 points · Estimate: 2 months
How to do this
Why it matters
Content that appears only after JavaScript runs can be absent for a crawler that does not run it. The page looks correct to a person and empty to a retrieval system.
How to do it
- Test what a crawler sees. Fetch the page with JavaScript disabled and read the raw HTML.
- Move the primary content into server-rendered HTML. Server side rendering or static generation both work.
- Keep JavaScript for interaction, not for the words.
- Check the hidden paths: tabs, accordions, infinite scroll and modals often keep content out of the source.
- Add a check to the release process so this does not regress.
You are done when
The raw HTML contains the full text of the page.
Common mistake
Trusting that every crawler renders JavaScript. Some do, on their own schedule. Many do not.
-
45
Re-message the brand for entity consistency
One description, one boilerplate, corrected everywhere at once.
2 points · Estimate: 3 months
How to do this
Why it matters
Every old profile, press mention and partner listing that describes you differently pulls the brand profile apart. At scale, fixing it is a communications project, not a website edit.
How to do it
- Agree one description, one boilerplate and one set of category words. Get sign-off once.
- Audit every external surface: press releases, partner pages, app stores, directories, conference bios, old campaign sites.
- Correct them in one coordinated pass. A slow rollout means both versions circulate together.
- Give every team the canonical block and one rule: no rewording.
- Retire old microsites and campaign domains rather than leaving them up.
You are done when
Ten independent sources describe the company in the same words.
Common mistake
Refreshing the messaging and leaving the archive. Old pages outrank new ones for years.
-
46
Build an in-house AI visibility function
Report share of model to leadership beside pipeline and revenue.
2 points · Estimate: 3 months
How to do this
Why it matters
Maturity means share of model as a standing metric: how often you appear against competitors across a full prompt set, reported like any other channel.
How to do it
- Define the prompt set. Cover the whole buying journey and hold it stable so the numbers compare.
- Measure monthly across the engines your buyers use. Record the competitor set every time.
- Report share of model beside pipeline and revenue, to the same audience, on the same cadence.
- Give the number an owner who can change the content, not only report on it.
- Set a target and review it quarterly.
You are done when
Leadership sees share of model in the standing report.
Common mistake
Buying a tool and naming no owner. The tool produces the number. Someone has to act on it.
-
47
Reorganize marketing around AI search
Redraw who owns content, PR, product marketing and analytics.
2 points · Estimate: 6 months
How to do this
Why it matters
A new job title on the old team does not fix this. Content, PR, product marketing and analytics each own part of the answer, and none of them owns the outcome.
How to do it
- Name one owner for AI visibility, with authority over content, PR and technical fixes.
- Move digital PR next to content. Third-party citation is now a content outcome.
- Give analytics a standing job on AI referral measurement, not a project.
- Change the incentives. A team paid on keyword rankings will keep producing keyword pages.
- Retrain writers on answer-first structure before you scale output.
You are done when
One person can answer for the score and can change the things that move it.
Common mistake
Redrawing the reporting lines and leaving the targets alone. People follow the target.
-
48
Renegotiate how your content is distributed
Rewrite syndication, partnership and licensing terms around citation value.
2 points · Estimate: 3 months
How to do this
Why it matters
If third-party placements carry more weight than owned content, your syndication, partnership and licensing terms are quietly working against you.
How to do it
- Review every agreement. Who can republish your content, under whose name, with what attribution and what link.
- Require attribution and a canonical or citation link in all new terms.
- Prefer partners whose pages already appear in AI answers. Test before you sign.
- Reconsider exclusivity. An exclusive that parks your best material on a low-visibility site costs you the citation pool.
- Renew on a shorter cycle so terms can follow the market.
You are done when
Your agreements name attribution and linking as terms, not as courtesies.
Common mistake
Syndicating with no attribution rules. The partner gets the citation and you get nothing.
-
49
Relaunch the site so agents can act, not just read
Let an agent book, quote, compare or buy on a buyer's behalf.
2 points · Estimate: 9 months
How to do this
Why it matters
The step past being described is being used. An agent that can book, quote, compare or buy on a buyer's behalf completes the task on your site instead of a competitor's.
How to do it
- Pick the transactions that matter: book a demo, check availability, build a quote, compare plans, reorder.
- Expose each one as an API with clear inputs, outputs and errors an agent can act on.
- Write the documentation as part of the product. An undocumented API gets no adoption.
- Make calls idempotent and states explicit, so a retry cannot double-book or double-charge.
- Keep a human path for everything. Agent access is an addition, not a replacement.
You are done when
An agent can complete a real transaction with you from start to finish.
Common mistake
Publishing an API with no documentation and no authentication story. Neither agents nor developers will touch it.
-
50
Publish a product feed to the agentic commerce protocols
Let an agent find, compare and buy your product inside the assistant.
2 points · Estimate: 3 months
How to do this
Why it matters
Assistants now complete purchases in the conversation. Inclusion runs on a product feed and a checkout interface that follow a protocol, not on how your website looks.
How to do it
- Decide which surfaces matter. The Agentic Commerce Protocol covers ChatGPT. Google's Universal Commerce Protocol covers Search and Gemini.
- Apply through the merchant program for each surface. A hosted store platform usually has a switch for this.
- Serve a complete product feed: identifiers, price, availability, shipping and return terms, refreshed daily.
- On a custom stack, build the checkout endpoints and the payment token flow the protocol defines.
- Keep stock and price live. A wrong price fails the order and costs you standing on that surface.
You are done when
An assistant can show your product, its true price and its real stock, and complete an order.
Common mistake
Treating it as a feed export. It is a live commerce interface with an uptime expectation.
Get your scorecard
Your scorecard is a one-page report. Keep it, share it with your team, or use it to plan the work.
- Your score out of 100 and your grade
- Your points in each of the four phases
- Every answer you gave, check by check
- The steps for your eight highest-return fixes
We do not rent or sell your details. See our privacy policy.
Your scorecard is ready
Your score is 0 out of 100: Invisible.
The report saves as an HTML file. Open it in any browser and print it to PDF.
Want us to do the fixes?
AGEO Agency does this work for small businesses and nonprofits. Start with a free AI visibility check.
Get my free AI visibility checkQuestions about the audit
Read the full method in AI Search Visibility Audit: 50 Checks in Four Phases.
This audit is adapted from the AI Search Visibility Audit by Victor Ramayrat on AgentLedGTM.com. Effort estimates are planning judgments, not measured figures.
How is the AI Search Visibility Audit scored?
The audit has 50 checks. Each check is worth 2 points, so the maximum score is 100. Yes scores 2 points, Partly scores 1 point and No scores 0 points. The score maps to five grades: Invisible, Trace, Findable, Quotable and Cited by default.
What should I fix first?
Start with Phase 1. Its 13 checks are settings, files and profiles, not content. One person can clear them in a working day. The fastest check takes about 15 minutes: remove a robots.txt rule that blocks AI crawlers.
Do I need a paid tool to measure AI visibility?
Not to start. Write 10 to 20 questions in the words your customers use. Ask them each month in ChatGPT, Perplexity, Gemini and Claude, in a fresh session. Record if you were named, if you were linked, who else was named and which pages were cited.
How often should I run the audit?
Run it each quarter. AI answer engines favor recent content, and several checks are ongoing work. A score that you earned once does not stay earned.
What is in the scorecard?
The scorecard is a one-page report. It shows your score, your grade, your points in each phase, every answer you gave, and the steps for your eight highest-return fixes. You can download it or save it as a PDF.