AI search
AI search: why your business is not in the answer, and what actually moves it.
Yes, we work on whether a business shows up when someone asks an AI assistant for a supplier like you. It is a real question with a mostly unglamorous answer: those answers are assembled largely from sources that sit outside your website, and being readable by the systems is not the same as being chosen by them.
We measure it per system, we publish the run counts, and we do not sell a score.
The situation this is usually bought to fix
Someone typed the company name into ChatGPT and found nothing useful. Or asked it to recommend a supplier of what you sell and got a list you were not on. Or a free tool reported that you are absent from AI answers, without saying why or what to do about it.
That is a reasonable thing to be concerned about, and it is worth sizing before spending anything on it. Discovery through AI assistants is real and it is growing. It is also, on everything we have seen, still much smaller than ordinary search.
We will not put a share figure on it. The published numbers come from vendor panels nobody outside those vendors can audit, and a number we cannot stand behind is worse than no number at all. If your enquiries have to improve this quarter, this is not the work that does it.
What the work involves
Check access first, then stop treating it as the answer.
Whether the retrieval crawlers can reach and read your pages. It takes an afternoon, its absence would be fatal, and its presence proves nothing.
Entity accuracy.
Whether the basic facts about your business — name, what you do, where you are, how to reach you — are stated consistently on your own site, in your structured data, and everywhere else that describes you. Branded answers are assembled from those sources, and inconsistency is what produces two systems describing the same company two different ways.
Source authority, which is mostly off your site.
The lists, directories, review platforms and third-party pages those answers are actually built from. This is the substantial part of the work and it is not publishing.
Citable pages on your own site.
Content that answers a question precisely enough to be quoted, with the claims sourced and dated.
Measurement, per system.
Repeated prompts, recorded per system, with the run counts, the date, the location, the language and whether the session was signed in — because all of those change the answer.
Training crawlers
Govern whether your pages may be used to train a model. Blocking one does not remove you from an answer.
- GPTBot
- ClaudeBot
- Google-Extended
- CCBot
Retrieval crawlers
Govern whether your pages can be fetched and read when a system is assembling an answer now. This is the access that matters, and having it proves nothing on its own.
- OAI-SearchBot
- Claude-SearchBot
- PerplexityBot
The tokens above are the ones named in primary vendor documentation. Which crawlers a business should allow is a decision with consequences in both directions, and it is one we set out rather than take on anyone’s behalf.
We measured ourselves first, and published the result
Across 28 unbranded runs, we appeared zero times.
On four systems, on one day, from Pune, in English, signed out. In 8 branded runs — where the systems were given our name — 7 described us substantially accurately.
Directory and listing sources appeared in 15 of those 28 unbranded runs. In one of them, ChatGPT stated in plain text that it had chosen the brands it named by reading two directories. We are listed on neither.
Our own site is fully readable by the retrieval crawlers, our robots.txt blocks none of them, and the result was still zero. Access is necessary and it is not sufficient — including for us, which is the part of this most often sold the other way round.
| Layer | The question | Our own result |
|---|---|---|
| Retrieval | Can a system find and read the site at all? | Works — full access, served correctly |
| Discovery | Does the business appear when the buyer does not name it? | Zero, in 28 unbranded runs |
| Recommendation | Is the business actually recommended? | Zero |
| Citation | Is the site used as a source in an unbranded answer? | Zero |
Two systems: not testable
Claude and Grok could not be tested at all. We hold no baseline for either. That is not a zero, and it is not a result — it is an absence of data, and we make no claim about either system.
28 unbranded runs on four systems, on one day, from Pune, in English, signed out. Counts, not rates. No confidence interval is claimed and none is available. Nothing on this table is averaged, blended or scored — the whole finding would be invisible under a single number, which is the argument for keeping the layers apart.
What we will not tell you about AI search
There is no AI visibility score here, and there will not be one.
We measured the same prompts repeatedly across systems: membership was often stable inside one system, and the order changed in every repeated prompt. A single blended number across systems destroys the only stable signal there is, and it would be precision we did not earn.
There is no “rank in ChatGPT”. There is no ranking to be had.
We have one measurement round, on one day. One measurement is not a trend, so we will not describe an improvement, a decline or a percentage change.
Two of the systems we wanted to test — Claude and Grok — could not be tested at all. We hold no baseline for either, so we make no claim about either.
- appearance in any system
- a mention, a citation or a recommendation
- a position — there is no ranking to be had
- an improvement, a decline or a percentage change
Nobody can promise those, and the mechanism is largely outside anyone’s site.
What we measure, and what we report
A set of prompts a real buyer would use, run repeatedly, per system, on a stated date, from a stated location, in a stated language, signed out unless there is a reason not to be. We record whether you appeared, in what form, and which sources the answer was built from.
You get frequency across runs, per system, never blended — and the sources, which are usually the actionable part.
What we cannot give you is a position, a score, or a comparison across systems that means anything.
One round, one day, one location, English, signed out. Two of the systems named above were not testable at all, and they are reported as untested rather than characterised.
The questions we get asked here
- “Everyone sells this now — what makes you different?”
- In the same 28 unbranded runs, two agencies in this city that sell exactly that did not appear either. That is one signed-out sample on one day and not a verdict on them, but it does suggest that selling the terminology is not the same as having the mechanism. What we offer instead is the diagnosis — absent from the answers, because those answers are built from sources you are not on — and the work that follows from it.
- “Ahrefs and Semrush give me a checker for free.”
- True, and they are good tools. They report the absence accurately. What they do not do is tell you which sources the answer was built from, or do the off-site work that changes it. That difference is where the work is — and it is worth being clear that the reporting part is available to you for nothing.
- “Can you get us into ChatGPT?”
- No, and anyone who says otherwise is selling you a guarantee they cannot honour. What we can do is fix the entity information, work on the sources those answers are assembled from, measure whether anything changes, and tell you honestly when it has not.
Across from this page: the organic search work. Above it: the whole system.
Where this sits in the rest of the work
One discovery surface, and today a small one.
It sits inside the same outcome as everything else we do: enquiries you can act on, measured. If a business is not being found in ordinary search either, this is not where we would start, and we will say so.