Evidence & Standards
What counts as evidence here, how it is collected, what it can support — and the claims it cannot be used to make, including some that would be commercially convenient.
The evidence is the recorded answers
Everything TendorAI reports traces back to an answer an AI assistant actually gave, recorded at the time it was given.
That is the whole basis. Not an estimate of how visible a business probably is, not a score derived from its website, and not a model of how assistants are believed to behave. A question was put to an assistant, an answer came back, and that answer was recorded and kept.
Keeping the answers matters more than it sounds. It means any figure we publish or send a client can be traced to the responses it was calculated from, and it means a claim can be checked rather than believed. A number with no recorded answers behind it is not evidence, however precise it looks.
How a measurement is designed
- A defined question set — the wording a buyer would actually use, agreed at the start and then held constant. Change the questions and the numbers move for reasons unrelated to the business.
- Repeated runs — the same question asked several times, because assistants do not return the same answer every time. A single run describes a moment, not a position.
- Named assistants — client measurement covers ChatGPT, Google AI Overviews and Perplexity. Our published studies name the assistants they used, which is not always all three.
- Recorded responses — kept as the evidence, not summarised and discarded.
- Citation capture — which sources appeared in each answer, recorded as cited.
- A stated window — every measurement carries the dates it was taken between, because assistants change without notice.
The same design applies whether the measurement is a client’s or one of our own published studies. The technical detail is set out on Methodology.
A cited source is not a cause
When an assistant shows the sources behind an answer, we record them as sources cited in that response. We do not describe them as sources that influenced the response, and the difference is not pedantry.
A citation tells you a source appeared alongside an answer. It does not tell you the answer was drawn from it, that removing it would change anything, or that getting your own domain into that list would alter what is said about you. Any of those might be true. None of them is established by the citation itself, and we have not run the experiment that would settle it.
So citation data is reported as what it is: a record of which sources appeared. Where we have a hypothesis about what it means, it is labelled as a hypothesis.
Observation, interpretation, and the gap between them
Four things get run together in this field, and we keep them apart in everything we publish or send.
- Observed results
- What the recorded answers contained. Facts about a sample, in a window, on named assistants.
- Hypotheses
- Things we think might explain the observation and intend to test. Labelled as such, and not used as the basis for a recommendation until they have been.
- Interpretations
- What we judge the observation means for a business. This is where our opinion lives, and it is presented as opinion so you can disagree with it while still relying on the measurement underneath.
- Unknowns
- What the measurement design cannot reach. Stated rather than left for the reader to notice.
The rule is simply that what was measured is said before what it is thought to mean.
What we will and will not say
What the evidence supports:
- A business appeared, or did not appear, in a recorded AI response
- A business was described in a particular way in a recorded response
- A particular domain was among the sources cited in a recorded response
- An answer occurred repeatedly across runs, or did not
- What changed between one measurement and the next
What it does not support, and what we therefore do not claim:
- That a page, a schema change or a content change caused an AI assistant to recommend a business
- That a cited source influenced the answer it appeared in
- That inclusion in, or claiming, a TendorAI profile causes AI recommendations or citations
- That a business will be recommended, cited or named in future
- That visibility will improve, in any stated period
Some of those would be easier to sell. The second list is the reason we can stand behind the first one.
Why a result has to be read with its design
A figure means nothing without the measurement that produced it. Our own published work is the clearest illustration.
In the August 2026 study we ran a fixed panel of 68 prompts across 17 UK cities against one assistant, Perplexity, ten repeats per prompt, between 18 July and 24 August 2026. Each of 1,214 SRA-regulated firms was assessed against the four prompts for its own city — 40 eligible answers per firm, 48,560 firm-answer observations. 1,003 firms, 83%, were never named once.
That number is real, and it is also narrow. It describes SRA-regulated solicitors, in those cities, on that panel, on one assistant, in that window. It is not a statement about all businesses, all industries, all assistants or “AI” in general, and we do not use it as one. Quoted without its design it would be misleading, which is why the design travels with it everywhere it appears.
The same study contains the most useful check on our own enthusiasm. A control group of firms we changed nothing about moved from 3.59% to 4.19% over five weeks — up 0.60 percentage points, for no reason we caused. Visibility drifts on its own, so any change worth reporting has to beat that drift before it means anything.
The standards, stated plainly
- Every published figure carries a date, a method and a sample size.
- Question sets are fixed before a measurement runs and are not changed to improve a result.
- Departures from a pre-registered plan are published in a deviations log, including corrections to our own errors.
- Datasets behind published research are made available, not described from behind a form.
- A finding about one sector stays a finding about that sector.
- Research URLs are permanent. Corrections and new editions are published without breaking existing citations.
- No record count is published until it has been verified against the database — we currently hold conflicting figures, so none is quoted.
- No outcome is guaranteed. We promise the correctness of the work, not what an assistant will say.
The published studies, their datasets and their deviation logs are at Research. The approach that produces them is on Our Approach.
Ask us how we would measure your business
Including what the measurement would not be able to tell you. That part is usually the more useful half of the conversation.