AI Visibility

What AI assistants see

It helps to separate five things that are easily run together, because they are not the same and only some of them are within anyone’s control.

• →The business itself — what it actually does, where, and for whom
• →The information available about the business — its own website and structured data, public registers, directories, reviews, and anything else a third party has published about it
• →The question being asked — its wording, the market it names, and which assistant it is put to
• →The response the assistant generates out of all of that
• →The sources cited in that response, where the assistant shows them

An assistant does not read your business. It reads the information available about your business and assembles an answer from it. Where those two have drifted apart — details that are out of date, services described in one place differently from another, a business well known locally but thinly described in public — the answer follows the information rather than the reality.

That is also why a business can be absent from an answer without having been rejected. It was never a candidate to be compared in the first place. And it is largely invisible from the inside: nothing appears in your analytics when an assistant names someone else.

How TendorAI measures AI Visibility

Measurement means putting a defined set of questions to AI assistants, repeating them, and recording what comes back.

The questions are agreed with the business and then fixed. Keeping them constant is the whole point: change the wording and the numbers move for reasons that have nothing to do with the business. Depending on what is being measured, a measurement run can involve:

• →A defined question set, agreed at the start and then held constant
• →Repeated runs of the same questions, so what is reported is a pattern rather than one answer on one day
• →The recorded AI responses themselves, kept as the evidence behind every figure
• →Comparison between assistants, where the question set is supported on more than one
• →Whether the business was mentioned, and how it was described when it was
• →Which sources were cited in those responses
• →Observations by market or location, where those are part of what was measured
• →Remeasurement of the same question set later, so one period can be read against another

Not every measurement includes every one of those. What a given run covers depends on what was agreed and on what the instrument supports for that business, and the report says which of them it is based on rather than implying the full list.

Two boundaries are worth stating plainly. A source appearing in a response is recorded as a source cited in that response — that is an observation, not evidence that the source caused the answer. And what a measurement covers is a question set, not the whole internet: it describes how assistants answered those questions, in that window.

The same discipline runs through TendorAI’s published work. Our August 2026 study put a fixed panel of questions to one assistant, Perplexity, across 17 UK cities and recorded, for each of 1,214 SRA-regulated solicitor firms, whether it was named in the answers it was eligible to appear in. 1,003 of them — 83% — were never named once. That is one sector, on one assistant, on one panel, in one window. It is not a claim about every industry, and it is the reason a business is measured before anything is said about it.

The methods and datasets behind that work are published in full under Research.

From evidence to action

Understand → Measure → Evidence → Find → Act → Remeasure → Compare → Report.

That is the customer-facing sequence, and it is a cycle rather than a project with an end date — because what assistants can find out about a business keeps changing as the business changes, the sources describing it change, and the platforms themselves change without notice.

1. 1

   Understand

   What the business sells, who buys it and which markets it works in. That is what decides which questions are worth putting to an assistant at all, because a question nobody asks tells you nothing.

2. 2

   Measure

   The agreed questions go to the assistants, repeated across several runs — one run describes a moment rather than a position.

3. 3

   Evidence

   Every answer is recorded and kept: what was said, whether the business was named, and which sources were cited in that answer.

4. 4

   Find

   The recorded answers are worked through to set out the gaps they actually support, each finding carrying the evidence behind it.

5. 5

   Act

   Priorities are agreed and the work is done — structured data, page and content changes, and corrections where public details are inconsistent. Nothing goes live without approval, and every change is recorded with its date.

6. 6

   Remeasure

   The same questions go back to the same assistants, unchanged. Changing the question set would move the numbers for reasons that have nothing to do with the business.

7. 7

   Compare

   The new measurement is set against the baseline: what moved, what did not, and by how much — reported as an observation, not as proof that the work caused it.

8. 8

   Report

   A written record of the measurement, the recorded answers behind it, what was changed and when, and how this cycle compares with the last. Then it repeats.

Each stage is set out at greater length on How It Works, which is the page that owns the process in full. Not every engagement runs the whole cycle: some stop at the evidence, some go as far as the findings, and some run the loop continuously. Which is which is set out under Products.

What AI Visibility does not mean

The term is new enough to be used loosely, so it is worth being explicit about what is not being claimed.

It is not a guarantee of recommendation.

No supplier controls a third-party AI platform, and those platforms change without notice. Measurement tells you what was said. It does not commit an assistant to saying it again.

It is not a ranking score we invented.

What is reported is what the recorded answers contained — whether a business was named, how often, how it was described, and which sources were cited. There is no league table behind it.

It is not proof that a particular source caused an answer.

A source appearing in an answer is evidence that it was cited. It is not evidence that it produced the answer, and we do not report it as though it were.

It is not SEO with a different label.

The two overlap and both are worth doing. They are not the same question, and this page sets out the difference above.

It does not mean TendorAI controls what an assistant says.

What is promised is the correctness of the work — that the measurement is real, that every figure carries a date and a method, and that the same questions are asked again afterwards so you can see what happened.

There is evidence behind that first point, and it is our own. In the August 2026 study a control group of firms we changed nothing about moved by 0.60 percentage points across five weeks — from 3.59% to 4.19% — for no reason we caused. Visibility drifts on its own. Any change worth reporting has to be read against that drift, and anyone promising a position in advance is either not measuring a control group or not mentioning it.