How Labs works

The SME Visibility Panel

The panel is 240 real UK small businesses across 9 sectors and 18 cities: accountants, builders, physios, restaurants, shops, agencies and more. Every one was found in a public place, with the source recorded. Nothing behind a login.

Each collection cycle puts the questions those businesses' customers might genuinely ask (“who are the best accountants in Leeds?”) to the AI engines people use, and records which businesses appear. Questions are generated from the sector and city, never from a business's name. Every question is asked three times, because AI answers change between asks. Collection began in August 2026 and runs in cycles of about four weeks of continuous sampling.

Where the bias is

Panel businesses came from public directories and lists, which is exactly what AI tools read. So the panel's numbers likely read rosier than the true national picture. The Index describes publicly listed UK small businesses, never all UK businesses, and says so wherever a figure appears.

The question set is locked and versioned

A reworded question becomes a new version and is never compared against the old one. The question set publishes in full alongside the first edition, with the cycles it produced.

Which cuts publish was written down first

Before anyone looked at a single aggregate, we recorded exactly which analyses will publish and which won't, dated in the method note. The discipline is the one medical trials use against cherry-picking. The first edition's questions:

• Which sectors are least visible? How often a sector's businesses get named at all when a customer asks who to use, by sector and by size of business.
• What gets recommended instead? When no real local firm is named, what fills the answer: directories, national chains, other independents, or press coverage.
• Do the AI tools agree? ChatGPT, Gemini, Perplexity and Google's AI Overviews, answering the same question on the same day, side by side.
• How much does the answer change between asks? The same question three times, and how often the businesses named change. This is why we report rates and not rankings.
• Discovery or validation? Whether businesses surface more when someone asks “who's the best?” or “is this firm any good?”

Why city-level results wait

Around thirteen businesses per city is too thin to defend. Cities wait until the panel is wider rather than publishing thin.

Two complete cycles before an edition

Edition 1 publishes when two complete collection cycles are in, and not before. We are deliberately not giving it a date: a date would be a promise about our schedule, and the only promise worth making here is about the evidence. The full method publishes at the Beyond the Beige summit in January 2027, before any results, so it can be challenged while it can still be improved.

The three kinds of study

• Surveys. A research company asks real business owners a set of questions on our behalf. We never contact anyone ourselves; respondents come from the research company's own opted-in pool.
• Experiments. We publish carefully matched pieces of content, identical except for the one thing we're testing, then track which version the AI tools cite, week after week.
• Analyses. We look for patterns in data we already hold, always anonymised and aggregated. No business or person is identifiable in anything we publish.

The rules every study follows

• Lock the question first. We record what we're testing and how we'll analyse it before looking at the results, so the data can't rewrite the question.
• Repeat the measurement. AI answers vary. Every number is measured across repeats, and we show how many.
• Publish the nulls. A study that finds no meaningful relationship, or contradicts what we expected, publishes with the same prominence as one that flatters us.
• Let people check. Methods are stated in enough detail to re-run them, and where we can do so responsibly, the raw data is downloadable from the study.

What we won't publish

• Any panel business, named. Not in an edition, not in a chart, not in a reply to a comment. They never agreed to be studied, so they are never identifiable, and we never contact them. We observe nothing a customer couldn't: public AI answers to ordinary customer questions.
• Any rank. Not “#3 in ChatGPT”, not “position 2”. We measure where a business appears in an answer and keep that internal, because publishing it would be a ranking by another name.
• Any single-sample claim, or any edition built on fewer than two complete collection cycles.
• Anything early. No teasers, no “early data suggests”. One leaked figure would undo the point of writing the analyses down in advance.

The raw data

Where a study's data can be shared without identifying anyone, it publishes with the study as a download. Panel data publishes as aggregates by sector and question type, with sample sizes, never as a list of businesses and what was said about each.

How we measure the product side is documented separately on How we measure.