TendorAITendorAI

Data & Evidence

The data behind our published research, available in full and ungated — with what each file covers, and what it cannot be used to show.

Why this is public

A finding you cannot check is an assertion. Publishing the panel and the classification list is what makes the studies arguable.

Everything below is published in full, with no registration and no form. That is deliberate: gated research cannot be cited, and research that cannot be cited cannot be checked by anyone who disagrees with it.

Reproduction is permitted with attribution to TendorAI and a link to the study the data belongs to. Figures, findings and charts may be quoted in full.

Published datasets

Three files. The prompt panel is unchanged between the July and August 2026 waves, so these serve both published reports rather than one each.

The full prompt panel

CSV · 68 rows
What it contains.
Every prompt used in both published studies, with its identifier, the city it covers, the practice area where one applies, and the exact wording put to the assistant.
Population.
Four prompt types per city across 17 UK cities — best-in-city, purchase intent, reputation and practice-area specialism. The seven practice areas apply to the 17 specialism prompts, not to the whole panel.
Window.
Fixed before collection and unchanged between the July and August 2026 waves, which is what makes the two comparable.
Limits.
A panel is a sample of the questions people ask, not all of them. Results drawn from it describe these questions and no others.
What it contains.
Each measured city and the number of SRA-regulated firms in the panel for it.
Population.
SRA-regulated solicitor firms in 17 UK cities. The August 2026 study assessed 1,214 of them, each against the four prompts for its own city.
Window.
The panel composition used for both published waves.
Limits.
This is a panel of solicitor firms in selected cities. It is not a sample of UK businesses, and nothing drawn from it describes businesses generally.
What it contains.
Each cited domain and the bucket it was classified into — directory, firm website, and the other categories used in the July analysis.
Population.
The domains appearing among the 12,279 citations recorded in the July 2026 study.
Window.
July 2026 collection.
Limits.
Classification is a judgement. It is published so the judgement can be checked and disagreed with, rather than presented as a fact about each domain.

The recorded answers themselves are not published as a file. The figures calculated from them are reported in each study, with the counts they were calculated from.

Which study, which assistants, which window

A result belongs to the assistants it was collected on. Both published studies name theirs.

The UK AI Visibility Report for Solicitors, July 2026 — TAI-R-2026-001
Two engines, Perplexity and ChatGPT. 1,360 answer runs on the 68-prompt panel across 17 cities. 12,279 citations recorded and classified. A single collection window in July 2026. The study states in its own limitations that Google AI Overviews reaches more UK users than either engine covered and is not represented in it.
Most UK Solicitors Are Never Recommended by AI, August 2026 — TAI-R-2026-002
One engine, Perplexity. The same 68-prompt panel across the same 17 cities, ten repeats per prompt, between 18 July and 24 August 2026. 1,214 SRA-regulated firms, 40 eligible answers each, 48,560 firm-answer observations. 1,003 firms — 83% — were never named once. A control group changed in no way moved from 3.59% to 4.19%, up 0.60 percentage points.

Both cover SRA-regulated solicitor firms in 17 UK cities. Neither is a sample of UK businesses, and neither is evidence about any other sector. Solicitors are the vertical we researched first, not the limit of who TendorAI is for.

Deviation logs

Every departure from a pre-registered plan is recorded and published, including corrections to our own errors. A study whose deviations are hidden is not reproducible, whatever its data says.

Published research and client measurement are not the same thing

They share a method but not a purpose, and the data on this page belongs entirely to the first.

Published research
Our own studies of a population — currently SRA-regulated solicitor firms. Designed to be citable, published ungated with the panel and classification list, and accompanied by a deviations log. Everything on this page.
Client measurement
A measurement of one business, on a question set agreed with that business, covering ChatGPT, Google AI Overviews and Perplexity. It belongs to the client and is not published. No client’s results appear in our research, and no research figure is presented as a client result.

Where the two meet is the method, which is set out once on Methodology and applies to both.

Evidence, and what we make of it

The files on this page are evidence: prompts that were run, cities that were covered, domains that were classified. They contain no conclusions.

The studies contain both, kept apart. What was observed is stated first, as fact about a sample in a window. What we think it means is labelled as interpretation. Where we have a hypothesis we call it a hypothesis, and where the design cannot reach something we say so rather than leaving the reader to notice.

  • A source appearing in an answer is recorded as cited in that answer — not as having influenced it.
  • A movement between waves is an observation. It is not proof that anything we or anyone else did caused it.
  • A finding about solicitors in 17 cities stays a finding about solicitors in 17 cities.
  • A figure quoted without its assistant, population and window is being quoted wrongly.

The standards behind that discipline are on Evidence & Standards.

These URLs are permanent

Published research and dataset URLs do not change. Corrections, new editions and methodology updates are published without breaking existing citations, because a citation that rots is worse than no citation.

If a figure is corrected, the correction is dated, the reason is given, and the earlier figure is restated alongside the new one so that editions stay comparable.

Check the work

The studies, the method and the data are all public. If you think we have got something wrong, the material to show it is on this page.