AI search visibility is whether AI Overviews, AI Mode, ChatGPT, Perplexity, Gemini and Claude mention your site when someone asks a question you should be the answer to, and whether they cite you when they do. It is a different measurement from ranking, no dashboard reports it, and most sites have never checked it.
This is the hub for that topic. It defines the thing, gives you a baseline procedure you can run this afternoon without buying anything, and says what moves the number and what does not.
The short answer
There is a second index now, and it does not publish a report. To find out where you stand, sample it: a fixed list of buyer questions, run against each engine, recorded by date, repeated. What moves the result, in order of how much it matters and how cheap it is, is crawler access, then whether your answers are extractable, then whether you are specific enough to be worth repeating. Everything else is detail.
Why this needs its own measurement
Search Console tells you about Google’s ranked results. It does not break out how often you were used in an AI Overview, and AI Mode traffic does not arrive as its own line either. Assistant traffic from ChatGPT or Perplexity shows up in analytics as a referral if the user clicks, and as nothing at all if they do not, which is the common case. Someone can read a paragraph you wrote, attributed to you, and leave you with no data that it happened.
So the visibility exists and the reporting does not. That gap is the whole reason this is a discipline rather than a setting.
It also explains why the field is full of confident claims. When nobody can see the number, nobody can check the claim. The honest response is to measure your own, badly at first, and keep the record.
The four surfaces, and why they are not one thing
People say “AI search” as if it were a single place. It is at least four, and they choose sources differently enough that being visible in one tells you little about the others.
AI Overviews. Google’s summary above the results, drawing on Google’s own index. Practically, ranking for the query is a strong prerequisite here, which makes this the surface most reachable through ordinary SEO. The posts on whether to optimize for AI Overviews and on the auto-expanding behaviour cover what changed and what it did to reporting. If you want the opposite outcome, keeping content out of AI Overviews is a legitimate choice with its own guide.
AI Mode. Google’s conversational surface. Same index underneath, different selection behaviour, and thinner reporting.
Assistants that browse. ChatGPT, Perplexity and Claude retrieve live pages at answer time through their own crawlers. Access here is a per-bot question rather than a general one, which is why the crawler posts are split by vendor: ChatGPT, Claude and Perplexity each run more than one user agent with different jobs, and blocking the wrong one has consequences people do not intend.
Model memory. What the model absorbed in training, with no retrieval involved. This is the surface you have least influence over and the one that changes slowest. Google-Extended is the control that touches it on Google’s side, and blocking it does something different from what most people assume.
The practical consequence: measure them separately. A site can be cited constantly by Perplexity and never appear in an AI Overview, and the fix for the second has nothing to do with the first.
How to baseline your own visibility
You do not need a tool to start. You need discipline about repeating the same thing.
1. Write the prompt list. Ten to twenty questions a real buyer would type, in their words, not yours. Include the category question (“best tool for X”), the problem question (“how do I fix Y”), the comparison (“X or Z for a small site”), and one that names you directly, which tests whether the engines know who you are at all. Write them down in a file. The list is the instrument, and changing it invalidates the comparison.
2. Run it against each engine, cold. Fresh session, no history, no personalisation, no follow-ups. Personalised context is the fastest way to measure yourself as more visible than you are.
3. Record two things per question. Were you mentioned, and were you cited with a link. They are not the same and they do not move together. A mention with no link is still worth having and is worth far less.
4. Record what was cited instead. This is the field people skip and the one that pays. If four sources came back and you were not among them, those four are the answer to what the engine considered good enough. Read them.
5. Date it, and keep it. One row per question per engine per run.
6. Repeat monthly. Same list, same wording, same cold conditions.
The long version of that procedure, with the prompt sheet broken down by question type, the five fields worth recording, and per-engine notes for ChatGPT, Gemini and AI Overviews, is in how to track brand mentions in AI search.
That is the whole procedure. What it lacks in sophistication it makes up for by being repeatable, which is the only property a baseline actually needs. When the list gets long or you want it run weekly across five engines, that is the point at which a tool earns its price, and not before.
Expect the first run to be humbling and the second to be inconsistent with it. Non-determinism is real here, and the way to survive it is a fixed list, a monthly cadence, and enough patience to ignore single-run movements.
What actually moves it
In the order I work through it, which is also cheapest first.
Access. Can the crawler behind that assistant fetch your pages at all. This is a yes or no question with a five-minute answer, and it is where a surprising share of zero scores come from: a plugin default, a firewall rule, a robots.txt inherited from a template. Nothing else on this list matters until this is clear. The vendor posts above list the actual user agents.
Extractability. Answer engines lift passages, not pages. A page that answers in its second sentence gets quoted; a page that answers in its ninth paragraph, after the preamble about how important the topic is, does not. Questions as headings, the answer immediately under the heading, one idea per paragraph. This is the single largest lever most sites have and it is a writing change, not a technical one.
Specificity. Generic text is not worth repeating, because a hundred other pages say it equally well and the model already knows it. What gets quoted is the sentence carrying a number, a date, a named constraint, a version, a limit. “Improve your content” is unquotable. “The export caps at 1,000 rows” is quotable. If a page contains nothing a machine could not have generated without you, it will not be cited, and it does not deserve to be.
Sourcing and identity. Being safe to cite means being attributable. A named author with a real record, claims tied to evidence, and consistent identity across the site and off it. This is slower than the others and it compounds.
Each of those is expanded, with the habits that implement it and the order to do them in, in how to improve brand visibility in AI search engines.
Then classic SEO. For the Google surfaces, ranking is close to a prerequisite, so indexing and technical health remain load bearing. A page that is not in the index cannot be summarised from it, which is where the ordinary indexing work meets this topic.
What does not move it
Publishing more. Volume is the standard answer and it is the wrong one here. Twenty thin pages give an engine twenty things not worth quoting. One page carrying a specific finding beats all of them, and a site that pivots to volume loses the one property that made it citable.
llms.txt, treated as a lever. It costs an hour, no engine is obliged to read it, and the honest framing is cheap and low risk rather than effective. What changed in v2 covers the detail. Write it, keep it in sync, do not expect it to carry weight.
Keyword density, and its new cousin. The idea that you can phrase your way into a model’s answer by repeating a term is the 2010 mistake wearing new clothes.
Anything sold as a guarantee. Nobody controls what an answer engine says. A vendor promising placement is either misunderstanding the system or counting on you to.
If someone is quoting you for this work, what generative engine optimization services actually do is the buying guide: what a real engagement contains, and the five questions worth asking before you sign.
My own number is zero, and that is the point
When this cluster was planned, the site scored zero mentions and zero cited pages across ChatGPT, AI Overviews, AI Mode and Gemini. Writing that down is deliberate. A specialist who claims to move AI visibility while never publishing his own baseline is asking you to take the one measurement that matters on trust.
So the number is public, the method above is the method being run, and what changes gets written up in the AI search posts as it happens rather than afterwards. Until the number moves, the service page describes the work and claims no outcomes. That is the correct order, and it is worth insisting on in a field this new.
Where to go next
If you have not checked whether the crawlers can reach you, start there: the ChatGPT, Claude and Perplexity posts each list the user agents and what blocking each one costs you. If you would rather not be summarised at all, keeping content out of AI Overviews covers the controls that exist. And if the pages are not in Google’s index in the first place, none of this is your bottleneck yet: why pages do not show on Google comes first.
If you want the baseline run against your site rather than a procedure to run it yourself, that is AI search optimization, and the first conversation costs nothing.
Before any of this is worth measuring, confirm the crawl access underneath it. Why ChatGPT cannot see my website walks the four OpenAI user agents and the three gates in order.
Quick answers
Is AI search visibility the same as ranking in Google?
Related but not the same. Ranking well helps, because several answer engines retrieve from a conventional index before they write anything, but it is not sufficient and it is not always necessary. A page can rank third and never be quoted because its answer is buried under four paragraphs of preamble, and a page can be quoted while ranking modestly because it states the thing directly and is safe to attribute. Rankings measure position in a list. AI visibility measures whether a machine chose to repeat you.
How do I measure AI search visibility when there is no report for it?
By sampling, deliberately and on a schedule. Write a fixed list of the questions a buyer would actually ask, run the same list against each engine, and record two things per run: whether you were mentioned and whether you were cited with a link. Keep the list, the wording and the date. The absolute numbers mean little on their own, so the value comes from running the identical list again later and comparing. A baseline you can repeat beats a sophisticated one you cannot.
How often should I re-measure?
Monthly is enough for most sites and weekly is mostly noise. These systems are not deterministic, so the same question can return different sources on consecutive days without anything about your site changing. Measuring too often turns that variance into a story. Leave enough time between runs that a real change has room to show up above the wobble.
Do AI visibility tools tell me anything I cannot check myself?
They mostly automate the sampling above, which is genuinely useful once the prompt list gets long or you want it run across several engines every week. What none of them can do is tell you why you were left out, and that is the part the work turns on. Start by running the list by hand a couple of times. You will learn more from reading the answers that excluded you than from a chart of how often they did.
Does being blocked to AI crawlers show up as low visibility?
Yes, and it is the first thing to rule out because it is the cheapest to fix. Sites regularly block the crawlers behind the assistants they want to be recommended by, usually by accident: a plugin default, a security rule, a robots.txt copied from somewhere else. Zero mentions with a blocked crawler is not an optimization problem, it is an access problem, and no amount of rewriting fixes it.