There are two conversations happening about AI search, and they are almost unrelated.
The first is a real, measurable question: when buyers ask an AI assistant about your category, what does it say, which sources does it use, and is what it says about you correct? That is answerable, repeatable and often urgent.
The second is a market of guarantees — "rank in ChatGPT", "our schema makes AI recommend you", "llms.txt unlocks AI visibility". Those are not supported by anything you can verify, and Google's own guidance contradicts several of them directly.
This piece is about how to work on the first one without buying the second.
Start with measurement, because it is the part that is real
A defensible AI visibility measurement has four properties: a fixed prompt set, multiple surfaces, several recorded dimensions, and repetition on a schedule.
Build a prompt set of five types. Eight questions each is a workable size:
- Discovery — "what are the best tools for [job to be done]"
- Category — "best [category] software for [segment]"
- Comparison — "[you] vs [competitor]", "[competitor] alternatives"
- Recommendation — "who should I use for [problem]"
- Brand — "what does [you] do", "is [you] any good"
Freeze the wording. This is the part that makes it measurement instead of anecdote. These systems vary between sessions, between users and between model versions. Only identical prompts, run repeatedly, produce a comparable rate.
Record six dimensions per answer, not one:
- Were you mentioned?
- Were you recommended?
- Were you cited — was your own domain a source?
- Which sources were actually used?
- Which competitors appeared?
- Was anything said about you factually wrong?
Run it monthly and report rates. A single check with a screenshot is not a baseline.
The most valuable finding is usually the least glamorous
In practice, the dimension that produces immediate action is factual accuracy.
Outdated pricing. A service scope you dropped two years ago. An integration you never had. A founder who left. These circulate because some third-party source said so once and nothing since has contradicted it clearly enough.
You can fix that in days, and it is actively costing you something today, with buyers forming a wrong impression before they ever reach your site. It is a better first project than anything strategic.
What genuinely appears to influence how you are described
None of this is a lever, and all of it is ordinary good practice:
State your facts plainly, in text. What you do, who it is for, which markets, what it costs, what it integrates with, who works there. Positioning copy is not a fact — "built for teams who move fast" tells a machine nothing. If a reader would have to infer it, so would anything else.
Make sure those facts survive without JavaScript. We regularly find pricing pages that render tiers from an API, so the raw HTML response contains no prices at all. The price that then circulates is whichever one a directory published two years ago.
Be consistent across sources. Website, LinkedIn, directories, profiles, press. Three contradictory descriptions of your product make all three less usable as a source.
Publish something worth citing. Original data, research, or a worked example. If everything on your site is undifferentiated marketing copy, these systems will reach for third parties — and they do.
Keep technical access working. The pages that explain your business need to
be crawlable and renderable. Check robots.txt for accidental blocks of
relevant crawlers.
Notice that every one of those also improves conventional search and the experience of a human reading the page. That is not a coincidence: Google's guidance treats AI search optimisation as part of SEO rather than as a separate discipline.
What nobody can sell you
A guaranteed mention or recommendation. There is no ranking lever, no submission endpoint, no paid inclusion for organic AI answers. A guarantee here means either the vendor does not understand the system or they are relying on you not to check.
Schema as a recommendation mechanism. Valid structured data is worth having where it genuinely describes a page. It helps machines parse content. Google's guidance does not require AI-specific schema for its AI features, and no markup makes a system recommend you.
llms.txt as a visibility strategy. Google has said it does not require llms.txt for its AI features. Publishing one costs nothing and some tools may read it. Treating it as an AI visibility strategy is not supported by anything measurable.
Manufactured mentions. Bought reviews, planted community posts, fake discussion threads. Beyond the obvious problems, Google's spam policies address scaled manipulation directly, and the reputational downside is asymmetric: it works until someone notices, and then it defines you.
How to report it without overclaiming
Report the rate change on the frozen prompt set, and separate two things carefully: what you changed, and what moved. Those are different columns.
Some months nothing will move. Say so. A report that only ever contains improvements is not a measurement system, it is marketing — and the person reading it will eventually work that out.
The one-paragraph version
Build a frozen prompt set across five question types, run it monthly across the surfaces you can access, and record six dimensions per answer. Fix every factual error you find, because that is both urgent and cheap. State your business facts plainly in server-rendered text and keep them identical everywhere. Publish something worth citing. Then report the rate change honestly, including the months it does not move — and do not buy a guarantee, because there is nothing behind it.
Want this done on your site instead of by you?
A Growth Diagnostic runs this analysis across search, technical, analytics, content, conversion and AI search, and comes back with what matters first.