The UK AI Recommendation Integrity Report, August 2026

Last Updated on September 5, 2026

The question this answers: after ChatGPT stopped leaning on Reddit in August 2026, who does AI cite for UK local recommendations, and how many of the businesses it names are real, trading firms?

In July, when we asked ChatGPT to recommend a trustworthy tradesperson in a UK town, Reddit was its favourite source. It appeared in 10 of 18 answers. More than the businesses’ own websites. Yell, Yelp and Facebook: zero.

In mid August, Reddit’s share of ChatGPT Search citations collapsed by 86% in four days.

Everyone is writing about what Reddit lost. Almost nobody is asking the question that matters to a customer in Preston: whatever fills that slot, are the businesses it names real? We have been checking AI recommendations against Companies House all summer. The honest answer is a spectrum. On our August panel, 5.2% of AI-recommended businesses failed every public check we could run, and that figure is a floor. Only 42% could be tied by name to a live registered company. The rest sit in between: probably trading, unverifiable by register.

This page reports what we measured, what we are re-measuring this week, and what a UK business should actually do about it.

What changed in ChatGPT citations in August

Credit where it is due: PromptWatch found this and published the data. Across its tracked prompt set, Reddit’s share of ChatGPT Search citations averaged 3.83% from 18 July to 7 August 2026, then 0.52% from 14 to 17 August, a drop of about 86%. On 8 August, ChatGPT’s background fan-out queries using the site: operator jumped from about 0.37% to 16.8%, and the average number of searches per response nearly doubled, from 1.08 to 1.83. The timing is not fully explained: the site: change came on 8 August, the citation cliff six days later, and PromptWatch itself said it could not yet rule out a measurement issue on its own side (Search Engine Journal, 19 August 2026). Google’s AI Overviews and AI Mode showed only gradual Reddit declines over the same period, roughly 11% and 30% relative, no cliff. OpenAI’s position, as reported on 20 August 2026, is that it does not target or set a fixed level of visibility for individual websites, and that Reddit continues to be cited. Nobody banned anything. What the logs show is a change in behaviour, not a policy announcement.

That is the last paragraph of this page about the aggregate chart. Everything below is UK measurement.

What we re-measured in the UK

Whito runs the same boring prompt at AI engines and writes down what comes back: “Recommend a trustworthy [sector] in [town].” One run per prompt, no retries, everything logged on the day.

We hold three relevant datasets. A 19 July citation log: 36 prompts across ChatGPT and Gemini, six sectors, three towns, every cited source recorded (method and log). The Q3 AI Visibility Index: 22 August, Google AI Mode, 20 towns, 8 sectors, 160 answers naming 734 businesses. And a 28 August verification pass checking each preserved recommendation against Companies House and the public record.

Update note, 29 August 2026. We are re-running the July ChatGPT panel, unchanged, against the post-change ChatGPT. No figure from that run appears on this page yet. When the results land, they will be added here with a dated note, per our corrections policy. Until then, this page makes no claim about which citation classes gained inside ChatGPT after 14 August.

One denominator note, because we said we would always make it: the Index recorded 160 answers and 734 names as collected; the archived logs preserve 159 answers and 728 names, and all recounts and verification on this page use the preserved figures.

Who got the citation slot Reddit lost

What we can already show is what the two worlds look like either side of the change.

ChatGPT, UK local prompts, 19 July 2026 (pre-change), 18 answers:

Google AI Mode, the same prompt shape at scale, 22 August 2026, 159 answers:

An own site or a Google surface appeared in 156 of 159 answers. Full source counts are in the citation sources study.

The two engines are not interchangeable, and we are not pretending AI Mode data answers a ChatGPT question. The point of the pair is the shape of the alternative: when an engine leans on domains rather than threads, the winners are the firm’s own site, Google surfaces, directories that bind a name to a town, and registers. That is the world ChatGPT’s new site: heavy searching looks more like, and the re-run will show how far it has moved.

Integrity: named businesses vs Companies House, scheme registers and the right town

A citation tells you where an answer came from. It does not tell you whether the plumber exists.

On 28 August 2026 we checked all 728 preserved recommendations from the 22 August panel against Companies House, scheme registers and the public record. The method distinguishes two standards, and the difference between them is most of the argument.

On the full-footprint standard, a recommendation passes if any public evidence confirms an active business trading in or serving that town: a register match, a scheme or platform profile, a live Google card, its own site. 38 of 728 failed everything. That is 5.2%, and because every close call was resolved towards verified, it is a floor. Of the 38: 21 matched only a dissolved, liquidated or dormant company, 15 had no company match and no public footprint at all, and 2 were names too generic to check.

On the strict register standard, a recommendation passes only if the name matches a live registered company in or serving the area. Just 305 of 728 managed that: 42%. The other 385 verified through a trading footprint or a national brand instead. Sole traders do not appear on Companies House, so this is not 58% fake. It is 58% of AI’s trustworthy names resting on signals no register underwrites.

Our July electrician study sat at the strict end and found the same shape: 93 of 182 AI-recommended electricians (51%) could not be verified as an active registered company in the recommended area, across ChatGPT, Gemini and Google AI Mode, and in 6 of 15 towns at least one engine recommended a firm whose registered company was dissolved or in liquidation. That is a prior finding, collected 9 July 2026. Whether the post-change ChatGPT does better on this exact test is what the current re-run measures, and the result will be added here.

So when someone quotes you a number for AI recommendations that are not real, ask which standard. The floor is 5.2%. The register-by-name figure is 58%. Both are true. They measure different things, and this page is the first place the two have been published side by side on the same dataset.

Sector split: trades vs professions

Integrity risk and platform dependence both split cleanly along one line: trades on one side, regulated professions on the other.

Unverifiable rates by sector, full-footprint standard, 28 August 2026:

Platform dependence, meaning the share of a sector’s town answers where a review platform or comparison portal sat behind the recommendation, 22 August 2026: roofers 100%, plumbers 95%, electricians 90%, cleaners 70%, estate agents 40%, accountants 25%, vets 15%, dentists 5%. Checkatrade alone was behind 50 of the 60 trade answers.

Read those two lists together and the story is uncomfortable for the trades: the sectors most dependent on platforms are also mid-table for unverifiable names, and the profession with the least platform dependence, dentistry, tops the failure table for a different reason: practice trading names that no register ties to a company. Nobody in this table gets to relax. And a detail for anyone who assumes regulators anchor these answers: across the 159 August answers, regulators and trade bodies were name-dropped 111 times but were the actual cited source only 16 times. Gas Safe was mentioned repeatedly. It was never once the cited source. Named, not read.

Why domain-level entity work now beats forum seeding

Here is the practical meaning of the August shift, argued from the logs rather than from a slogan.

ChatGPT is searching more, and more of its searches are scoped to named domains: site: fan-out went from 0.37% of queries to 16.8% in a day, and searches per response nearly doubled (PromptWatch, August 2026). A site: query cannot land on a rumour. It lands on a domain, and the domains that answer local trust questions are the firm’s own site, the scheme registers, and the platforms that already bind a legal name to a town. That is precisely the mix Google AI Mode already showed at scale in our August panel: own site 90%, directories 53%, registers 45%, forums 32%. On PromptWatch’s aggregate figures, Reddit’s share inside Google’s AI surfaces has been sliding all summer as well, gradually rather than off a cliff.

Forum seeding, the tactic of paying someone to say nice things about you in threads, was always rented visibility. Our July log shows how good the rent briefly was: Reddit in 10 of 18 ChatGPT answers. The August data shows the landlord repricing without notice. Reddit did not become worthless, and we are not claiming it did; real conversations there still matter and still get read. But a tactic whose entire yield depends on one engine’s weighting of one site’s threads is a tactic that can lose 86% of its distribution in four days, because in August it did.

Entity work is the opposite trade. A Companies House record that matches your trading name, a register entry that matches both, an owned domain that states name, place and proof: these are boring, cheap, and legible to every engine at once. When engines disagree about everything else, they agree on domains.

Why we are not recommending llms.txt

You will be sold llms.txt this month. Here is why we will not sell it to you, from server logs rather than opinion.

Ahrefs analysed 137,210 domains and found 97% of llms.txt files received zero requests; of the requests that did arrive at the other 3%, 96% came from bots, mostly SEO tools studying the files, with AI retrieval bots at 1.1%. No AI bot went looking for an llms.txt that did not exist. EZY Research put llms.txt on 83 sites and watched logs for 12 weeks to 19 July 2026: OpenAI’s crawler fetched robots.txt 3,990 times and llms.txt 7 times. Anthropic: 3,120 against 9. Perplexity: 775 against 0.

Our own audit of 120 UK small business sites found exactly one llms.txt written by a human being; 20 of the 21 files that existed were auto-generated by a platform or plugin. Which means agencies are billing for a file the answer bots almost never read, and the file most sites have was produced by a checkbox. Whito has an llms.txt. It cost nothing and it is honest about its own weight. If someone quotes you a price for one, that is the tell about the rest of their invoice. robots.txt is where AI access is actually decided today; the crawlers say so 3,990 to 7.

What a UK business should do this month

Free actions first. Structure before scale. Every item below is legible to a site: search.

1. Match your names. Your trading name, your Companies House record and your website footer should agree. 385 of the businesses AI called trustworthy in August verified only by footprint, not by name. If your register entry does not match your sign, you are one of them. Cost: nothing.

2. Check your register entries. Gas Safe, NICEIC, TrustMark, GDC, ICAEW, whichever applies. Same trading name, current address, live status. Engines name-dropped these registers 111 times in our panel; make sure what they find when they do read yours is you. Cost: nothing.

3. Say the basics on your own domain. Who you are, where you are, what you do, with your company number on the page. A business’s own website was the most cited source class in the August panel at 90%. Not a redesign. A page. Cost: nothing.

4. Keep your Google profile alive. Google surfaces fed 81% of answers. Reply to reviews, keep hours current. Cost: nothing.

5. Trades: your platform profile is your shopfront whether you like it or not. A platform sat behind 100% of roofing answers and Checkatrade behind 50 of 60 trade answers. If you are on one, make the listing match your legal identity. We are not telling you to pay anyone; we are telling you what the engines cited.

6. Do not buy forum seeding, and do not buy an llms.txt. The first lost 86% of its ChatGPT distribution in four days. The second gets read 7 times per 3,990. Spend the money on jobs 1 to 5, which cost nothing, and put the savings back in the van.

If your marketing spend this month does not survive contact with this list, that is the audit finding.

Dataset and how to cite this page

The row-level verification sheet (728 recommendations, coded per the published codebook) is available to journalists and researchers on request: hello@whito.co.uk. A downloadable CSV will be added to this page together with the ChatGPT re-run results.

External figures cited above: PromptWatch (August 2026) for the citation-share and fan-out data; Search Engine Journal, 19 August 2026, on the disputed timing; Yahoo Tech, 20 August 2026, for OpenAI’s position; Ahrefs llms.txt study (2026, data from May 2026); EZY Research llms.txt log study, 27 July 2026. Whito figures come from the studies linked in the body, all under /research/.

Common questions

Did ChatGPT ban or block Reddit?

No, and this page does not claim it. PromptWatch measured Reddit’s share of ChatGPT Search citations falling from 3.83% to 0.52% between early and mid August 2026. OpenAI says it does not set fixed visibility for individual sites and that Reddit is still cited.

How many AI-recommended UK businesses are not real?

It depends on the standard. On Whito’s August 2026 panel of 728 recommendations, 5.2% failed every public check (a floor), while only 42% matched a live registered company by name. The honest range for “not verifiable as an active registered company” in Whito’s 2026 studies runs from 5.2% (full footprint) to 51% (strict register standard, July 2026, three engines).

Which sector has the least reliable AI recommendations?

Dentists, in the August 2026 data: 11.2% of AI-recommended dental practices could not be verified as an active business trading in the town named, against 1.2% of accountants.

Should a UK business post on Reddit to get recommended by AI?

Not as a strategy. Reddit fed 10 of 18 UK ChatGPT recommendation answers in July 2026 and then lost 86% of its ChatGPT citation share in four days in August. Genuine participation is fine; paid seeding is a rented tactic priced by someone else.

Does llms.txt help AI find my business?

The server logs say almost never. Ahrefs found 97% of llms.txt files receive zero requests, and EZY Research logged OpenAI fetching robots.txt 3,990 times against 7 llms.txt fetches across 83 sites in 12 weeks.

What is the single free thing to do first?

Make your trading name, Companies House record and website footer agree. In August 2026 only 305 of 728 AI-recommended businesses could be tied by name to a live registered company.

Corrections and right of reply: if your business appears in the underlying dataset and you believe a fact is wrong, email hello@whito.co.uk and we will check it and correct it, with a dated note. Policy: whito.co.uk/corrections/.

Ethan Whitmore Co-founder and Analyst

Ethan Whitmore is a co-founder of Whito. He has spent over nine years in SEO and ecommerce inside global SaaS platforms, enterprise brands and payments companies. His work covers technical SEO, content strategy, conversion and video production.

See Full Bio