Which AI surfaces can actually be measured through an API, and which cannot

Every tool in this category reads AI answers through something, and what it reads through decides what the number means. The specification an engine wrote for this question asked for a verified four engine coverage rate, expressed as successful engine checks divided by attempted checks, with the test window and the exact plan disclosed. We do not hold that rate and this page says so rather than producing one. What we do hold is four measured limits from our own run, and each of them changes how a coverage claim should be read.

Limit one, the location parameters did nothing

The interface we read ChatGPT and Gemini through accepts four parameters that look like geography: gl, hl, location and country. It ignores all four. We measured that on 22 October 2026. The entire United States demographic in a 635 reading run was therefore produced by a market lock line written inside the prompt text.

What this means for a buyer. If a vendor says their readings are taken in your market, ask whether the market is set by a parameter or by prompt text, and ask them to show you a reading where the two differ. A geography field that has no effect is worse than no field, because it produces a confident label over an unlocked reading.

Limit two, a 4,000 character cap met a 21,711 character answer

Our pipeline has three steps. Ask the buying question. Feed the engine its own answer back and ask which sources it used. Then ask what a vendor would have to publish to be worth citing. The second step cannot run through the interface at all, because it needs the engine's own answer in context and the interface caps a prompt at 4,000 characters, while one batch answer in this run measured 21,711 characters.

So that step ran by hand, in the products, for 38 pipelines, and the third step ran through the interface standalone with the questions as context instead. A pipeline described as fully automated across all surfaces is describing something we could not do, and the character cap is the reason rather than effort or budget.

Limit three, 85 of 635 readings were not readings of the live web

Of 635 scored readings, 550 ran a live web search. The remaining 85 answered from training data. Both are legitimate objects of measurement and they are not the same object. A reading with no retrieval had no opportunity to cite anybody, so counting it in the denominator of a citation rate describes something other than citation. We keep the search flag per reading for exactly that reason.

The same split showed up in a separate consistency run on the same day, where ChatGPT searched in 66 of 66 readings and Gemini in 59 of 66. The detail is in the run to run measurement.

Limit four, the engine list you can actually verify is shorter than the engine list you can name

In our own audit the measurement arm covered two engines, ChatGPT and Gemini. Claude appears in the specification arm only, and only through the product interface by hand, which produced 6 of the 224 specifications. Perplexity and Google AI Overviews are not in this run at all.

That is why there is no four engine coverage rate on this page. Writing one would require 10 test prompts per engine per vendor, which is 40 checks for a single functional audit, plus successes over attempts reported per engine with failure and retry rates. We have not run it against anybody, including ourselves.

What we can report is what the pages publish, on the dates we read them. Our own product reads five surfaces, and a Claude check costs 3 credits rather than 1, read from our pricing page on 22 October 2026. Fetched by us the same day, tryprofound.com/pricing publishes three answer engines on its Trial tier, named as ChatGPT, Gemini and Google AI Overviews, with 50 unique prompts run daily for 7 days, one language, one region, and no history, no exports and no API; its Enterprise tier publishes up to 9 answer engines, all time history, CSV and JSON export, an API, SSO and SAML with SOC 2 compliance, and the price as Custom. There is no price figure in any currency anywhere on that page, so any dollar figure for that vendor in circulation did not come from it.

How to test a coverage claim yourself, in one afternoon

1. Pick one buyer question and five engines. Ask it by hand in each product interface, on the same morning, and save the answer text.
2. Ask the tool for the same question on the same day and compare the names, not the score.
3. Check the geography. Ask the same question with a market locked in the prompt text and without it, and see whether the tool's answer changes when its own region setting changes.
4. Count the engines you can confirm, and treat every other engine in the brochure as unverified until you have confirmed it.
5. Record whether each answer cited anything at all. That flag is what makes a citation rate honest.

What this page does not establish

One interface, read on 22 October 2026, in our own hands. A different data provider may honour the parameters this one ignored, and a cap may move. These are limits we measured, not properties of the engines themselves, and the right response to a vendor disagreeing is to ask them for their own dated measurement rather than to treat ours as the last word. We sell measurement software, so the gaps above are gaps in the thing we sell.

Our own pricing page renders no SSO, SAML, SOC 2, ISO, DPA or white label, checked by word boundary over 119,203 bytes. A loose search for sso on that page matches crossOrigin seven times and subprocessors twice, which is how a feature gets reported in both directions by people reading the same file. The grading criteria for a monitoring tool are in how to grade an AI monitoring tool.

Questions about measuring AI answers through an API

Is an API reading the same as what my buyer sees

Not reliably. In our run 85 of 635 readings had no live search behind them, and the four location parameters the interface accepted had no effect on the answer.

Can every engine be read programmatically

Our measurement arm covered two. Claude entered our run only through the product interface, by hand, and Perplexity and Google AI Overviews were not in it at all.

What should I ask a vendor about coverage

Successes over attempts per engine, with the window and the plan. Ask for the denominator. A list of logos is a brochure and a ratio is a measurement.

Why does the citation audit step matter

Because it is the step that separates what an engine cited from what it claims it cited, and in our run it could not run programmatically at all against a 4,000 character cap.

The same run produced two findings about the data it collected. Hostnames that were not sources at all are in two of the twenty most cited hosts in our citation data were not sources at all, and the host types that did hold the citation ground are in documentation outranked every vendor blog in our citation data, and what a B2B company does about it.