Search Console sees Google. Buyers ask elsewhere.

Open Search Console and find a page showing up in Google's AI features. Good. That is evidence that Google put a link to your page in an answer. It is not evidence that ChatGPT, Claude, Perplexity, or Gemini would recommend your business to the same person with the same question.

The confusion is understandable because both things are called AI visibility. They are still two measurements from two different systems. Search Console observes Google's own search surfaces. Prompt monitoring recreates questions on a chosen set of assistants. Neither substitutes for the other, and putting them in one score hides the question each one actually answers.

The Stack House makes the same useful distinction in its Spanish guide, Search Console now measures AI visibility. ChatGPT remains outside the report. It is worth reading for the Search Console workflow. This article is the missing next step: how to join that Google signal to a prompt measurement without pretending they are the same number.

Google's own documentation is deliberately modest. AI features such as AI Overviews and AI Mode use the same technical requirements, crawling systems and page-level fundamentals as normal Search. A page needs to be indexed and eligible to appear with a snippet before it has any chance of being surfaced there.Google's AI features documentation is the technical reference, not a GEO checklist with a secret extra step.

When a generative-AI report is available for your property, use it to answer a narrow but valuable question: which of our pages is Google choosing to expose inside its generative experiences? Look at page, country, device and date. Compare similar pages, not the entire site against itself. A category page and a definition page have different jobs and should not be expected to move together.

That is a better starting point than looking at a sitewide “AI score.” A page-level increase says a particular URL was selected more often by Google. It gives you a place to inspect: the direct answer, the table, the product detail, the source citations, and the competing pages that do the same job better.

What it cannot tell you

It cannot identify a buyer's original wording when Google does not expose it. It cannot separate every Google generative surface into the exact question-answer event you want. Most importantly, it cannot observe a recommendation made by a non-Google assistant. A page selected for an AI Overview may never be named by Claude. A product named first by ChatGPT may not have a link in a Google answer at all.

The instinct is to go to Analytics and look for referrals. That is useful, but it comes late in the chain. A person has to see an answer, decide your link is worth opening, and arrive with a referrer that survives. Most recommendations do not complete all three steps.

GA4's AI Assistant channel makes this cleaner for some sources, but its own rules are a source list, not a census of assistants. The list changes, and it does not turn a Google AI Overview click into a separate AI measurement. Read the live default channel group rules, then read The AI traffic GA4 cannot seefor the holes that remain.

Cloudflare's AI crawl-to-refer data makes the shape plain. An assistant can fetch many pages for every visit it sends. The fetch is evidence that a system reached your site. The referral is evidence that somebody clicked. Neither is evidence that the answer recommended you over a competitor. The live ratios on Cloudflare Radarare worth watching precisely because they stop you treating a small referral number as the whole interaction.

Start with the questions a buyer asks when they are close to a decision. Not your head keywords. Questions such as “what is a good alternative to X for a five-person team”, “which tool handles Y without Z”, and “what should I use if I need A and B together.” The words after “without” are often more useful than the category in front of them.

Ask a fixed set of those questions on the assistants you actually care about. Save the full response, every brand named, the order of those brands, and any source URLs. Then ask them again next week without tidying the prompt set halfway through. This does not measure the world. It measures a repeatable sample of the buyer conversations you chose, which is the honest unit of work.

• Name the prompt verbatim. A near-identical rewrite can produce a different answer.
• Record each assistant separately. ChatGPT being present says nothing about Claude.
• Keep the named brands and their order. Being named last is not the same result as being named first.
• Save citations and response text. A score without the answer that produced it cannot tell you what to change.
• Repeat before you claim movement. One answer is a sample, not a trend.

That last line is where most visibility reports stop being useful. Assistants vary their answers. A single prompt can name you once and omit you on the next run with no website change in between.Proving a change moved an AI answerlays out the before-and-after design: repeated runs, a fixed prompt set and a holdout set so you do not congratulate yourself for a fluctuation.

Most teams begin with the phrases that already sit in their SEO spreadsheet. That is understandable and usually wrong. A keyword tells you that someone typed a short fragment into Google. It does not tell you what they ask an assistant when they want help making a choice. “CRM software” is a market. “Which CRM should a two-person consultancy use if it needs email history but will not hire an admin?” is a decision.

Take the questions from places where decisions have already shown themselves. Sales call notes. Support conversations. The comparison searches that lead to your site. The reasons customers give when they choose a competitor. The conditions in a prospect's final email. You will get a list that is messier than a keyword export, and that is the point. Buyers are messy too.

Split that list into four jobs. Category prompts ask who belongs on a shortlist. Alternative prompts ask who replaces a named incumbent. Constraint prompts ask which product works under a specific limit. Implementation prompts ask how something actually connects, migrates or complies. A business can look strong in the first group and invisible in the fourth. One average hides that difference and sends the content team after the wrong pages.

Do not change the wording every time you learn something. Add a newly discovered question to a separate set and label the date it joined. The old set is how you compare the month. The new set is how you explore. Mixing them lets a dashboard call a wider sample a gain or a loss, when neither happened.

You do not need one dashboard to rule them all. You need a short review where every number keeps its label. Once a month is enough for most sites. The useful sheet has four rows.

• Google generative exposure: which pages Google showed in its AI search experiences, from Search Console.
• Organic search demand: impressions, clicks and page-level trends, so an AI feature does not get blamed for a broader demand change.
• Recommendation frequency: how often each selected assistant names you across the tracked prompt runs.
• Competitive evidence: which competitor was named instead, which page or source was cited, and what claim their answer fulfilled.

The rows influence each other, but they are not conversions of one another. High Search Console exposure with low ChatGPT presence means one thing. High ChatGPT presence with no Google generative exposure means another. A blended score can make both patterns look average, which is the one result that gives you nothing to do next.

A review that ends with a decision

Start with one page that gained or lost Google generative exposure. Read the page, then the nearby pages that did not move. Next, take one buyer prompt where you are absent and read the answer, including the citations. Put the two pieces of evidence beside each other. The point is not to find a correlation that sounds good in a slide. It is to choose a claim a page can make more directly, a comparison it can answer, or a source it needs to earn.

Write that decision down before opening a CMS. “Improve AI visibility” is not a task. “Add the implementation limits that made three assistants choose Competitor A for this migration question” is a task. It has a page, an owner, a before state and a question you can run again later. This is boring project management. It is also the distance between a monitoring habit and an expensive screenshot folder.

Do not start by adding an llms.txt file or a paragraph about being “AI-ready.” Start with the prompt you are losing and read the actual answer. The usual gap is embarrassingly concrete: the competitor has a comparison page, a clear eligibility statement, a table of trade-offs, a better documented integration, or a third-party source that says the thing your own site only implies.

The page that ranks first is not automatically the page an assistant names. Search results are lists. Answers are short selections. Ranking is not being namedgoes through the places the two diverge and the kind of page that can close the gap.

Make one change that answers one missing claim. Publish it. Leave the tracked prompts alone. Wait for the page to be crawled and indexed, then run the same sample again. If Google begins showing the page more often but the recommendation rate does not move, you learned something real: visibility in Google was not the limiting condition for that assistant and that prompt.

This one feels backwards until you remember that the assistants are not all using the same retrieval step. A model may name a product from its learned knowledge, from a different web index, from a source with no Google visibility in the market you tested, or from a response that does not include a link at all.

Do not use that as a reason to ignore Search Console. It is an invitation to split the work. Keep the product facts that made the assistant name you clear and current. Then fix Google's separate technical and content requirements: indexability, canonical URLs, snippets, page purpose and the pages competing in that search result. Google's own rules remain the baseline here.

This is the point of measuring both layers. One system can show a win while the other shows work to do. That is not a contradiction. It is a map.

If you have never measured this, do not begin with a category score or a 200-prompt spreadsheet. Pick one important page and a handful of buyer questions it ought to answer. Check the page in Search Console. Run those questions on the assistant coverage you care about. Read every answer. Write down what surprised you.

Beseen turns that first pass into a free report: a domain, a set of discovered searches on ChatGPT, Google positions beside them, and one rewrite for the worst gap. It is not proof that the rewrite worked. That needs the second check, which is the part the first report cannot do for you.

Start with a free report if you want the first prompt set and the page behind the most obvious gap. Then use the next check to find out whether the answer actually moved.

Can Search Console show whether ChatGPT recommends my business?

No. Search Console is Google's reporting product. Its AI data can show Google's own generative search exposure when that reporting is available for your property. It cannot observe ChatGPT, Claude, Perplexity or another assistant's answers. Run a stable set of prompts on the assistants you care about, then keep the results separate from Search Console data.

Does a Google AI Overview citation mean I will be recommended elsewhere?

No. It means Google chose to show your link in that experience. Other assistants may use different retrieval, different source selection, different model behavior and different answer formats. A citation is useful evidence, especially when it points to a page worth improving. It is not a passport that transfers to every other assistant.

How often should I measure AI visibility?

Weekly is usually enough to spot a sustained change without teaching yourself to react to every wobble. Run more repetitions when you are testing a specific claim, not because a dashboard made daily movement look important. Keep the prompt set, assistant selection and the way you record brand order stable across the comparison window. Otherwise a new number can look like progress when you only changed the measurement.