On 3 September 2026 fifty buyer questions were put to ChatGPT, Gemini, Perplexity and Google AI Mode from an anonymous browser session, one run per question, 193 answers in all. The questions are the ones a buyer types: what an agent readiness audit is, what it costs, who does it in Finland, how it differs from SEO. Claude was not measured, because its answers sit behind a login and the instrument runs logged out.
Two products share the words
Eleven of the fifty questions use the words agent readiness, or agenttivalmius in Finnish, without saying whether they mean a website or a company. Those eleven got 41 answers. Eighteen of the 41 described the readiness of an organisation to deploy AI agents: its data, its governance, whether its teams are prepared. Thirteen described what this site means by the words, whether a website and its APIs can be read and acted on by an agent. Six described both, and four could not be placed.
The split is not even across assistants. Gemini gave the organisational reading in eight of its eleven answers and Google AI Mode in six of nine. Perplexity was the only one that leaned the other way, five of eleven for the website reading, and it was also the one that named this site most often, five times against three for ChatGPT and one each for the other two. ChatGPT split evenly, three answers for each reading and three for both, with one unclear.
The words are the problem, not the assistants
Two Finnish questions in the set name the website outright or avoid the word agenttivalmius. All seven answers to those two read the question as being about the website. The assistants have no trouble with the concept once the question carries it. What drifts is the bare term, and it drifts toward the organisational meaning. The likely reason is what the assistants have read, because that meaning appears in far more published text, but this run does not measure that, and the control set is seven answers, so read it as the direction and not the proof.
That shows in who gets named. Thirty-five of the 41 answers named at least one provider, 112 different names between them, and almost none repeated. Both kinds of answer named a long tail of small consultancies and scanners with agent readiness in the brand, most of them once. The organisational answers also named large consultancies, and the Finnish questions a few Finnish agencies. Across all 193 answers this site was named in 27, and 23 of those carried a link to it.
What follows from it
For a buyer, one question settles it: ask the provider what the report reads. If the answer is the organisation, its data and its teams, that is a strategy engagement. If the answer is the website and the API, the files an agent fetches and the data it gets back, that is the audit this site sells, and the report should name the checks. Both are real work. They are not the same work, and a proposal that does not say which one it is has not said much.
For this site, the same finding turned into one sentence. The FAQ on the services page and the home page now says that the audit reads the website and its APIs and not the organisation's readiness to adopt agents, and llms.txt says the same in its summary, so an assistant that reads either gets the distinction in the first lines. Whether that moves the answers is what the next run of the same fifty questions will show, and it will be reported the same way.
Frequently asked
Is an agent readiness audit about the company or the website?
Both services exist under the same words. The audit on this site reads the website and its APIs with an independent scanner and reports per check. An organisational agent readiness assessment reads a company's data and governance and how prepared its teams are. Ask a provider which report they deliver.
Why do AI assistants describe the organisational version more often?
In a measurement of 41 answers on 3 September 2026, 18 described the organisational reading and 13 the website reading. When a question names the website, all seven answers in the control set read it that way, so the drift sits in the bare term and not in the assistants' ability to read the website meaning. Why the bare term drifts was not measured.
How was this measured?
Fifty buyer questions, four assistants, one anonymous run per question on 3 September 2026, 193 answers. Each answer was read for how it interpreted the term, whom it named and whether it named this site. Claude was not measured because it requires a login.
Corrected 2026-09-04. The version served for the first seven hours said Perplexity was the only assistant that named this site more than once and offered an unmeasured reason for the drift of the term. Perplexity named it most often, five times against three for ChatGPT, and the reason is now stated as unmeasured. The counts did not change.
Related
- What an agent-readiness audit is
- Choosing an agent-readiness audit
- Agent-readiness, AEO and GEO: how they relate
- Website agent readiness, measured on 567 company sites
Originally published at https://turva.dev/blog/what-ai-assistants-call-an-agent-readiness-audit
Top comments (3)
This is wrong-candidate resolution at the vocabulary level — the same failure mode as a symbol lookup with two same-named definitions. A bare term with two referents gets silently resolved to one interpretation, and the answer is well-formed about the wrong object. Your control set nails the boundary: assistants resolve fine when the question carries the referent, the drift lives in the bare term itself.
What your next run should also show: whether the llms.txt definition acts like an explicit disambiguator — the same fix my resolver took (return candidates or require explicit targeting instead of first-match). If the drift toward the organisational reading shrinks after your definition ships, that's evidence that machine-readable definitions compete with corpus frequency — which would be a genuinely useful finding for anyone maintaining llms.txt.
The sequence happens to be clean. That run was 3 September and the definition shipped on the 4th, so the before is on record and predates it, which was luck rather than design. Your control is already in the data as the seven answers to the two questions that carry the referent, and if those hold at seven while the bare-term answers move, the movement is in the term. One confound worth naming before the next run: llms.txt disambiguates only if something fetched it, and the instrument does not measure that, so a null result would not separate a weak definition from an unread one.
The confound has a cheap fix from your own toolkit: make the definition fetchable in a way that leaves a trace. Three options, from cheapest to strongest:
Without one of these, the next run's null result stays ambiguous — and the ambiguous null is exactly the absent answer your whole measurement series is built to eliminate.