# How visible are UK small businesses to AI? Our 2026 homepage study
> A measured convenience cohort of 301 UK-facing business sites: homepage structure, crawler policies, missing data and a transparent downloadable method.
- Page to cite: https://aeolookup.com/ai-for-small-firms/uk-business-ai-readiness-study-2026/
- Publisher: AEO Lookup (https://aeolookup.com/)
- Section: AI for small firms, Putting it to work
- Published 30 October 2026
- Key fact: We attempted 301 distinct business domains and measured 236 eligible homepages; 65 attempts were excluded. This is a convenience cohort, not a national estimate.
- Sources: AEO Lookup first-party collection receipts, 30 October 2026; scoring source e2afbef; observations fingerprint 289c2d00524776d2; Google Search Central AI features and your website; OpenAI crawler documentation; llmstxt.org proposal. Full fingerprints and scope are in the linked method download.
In our 2026 study, 155 of 236 successfully measured UK-facing business homepages (65.7%) contained JSON-LD structured data. These are technical observations of homepage HTML, not measurements of how often ChatGPT or another assistant cites those businesses.
We attempted 301 distinct websites, with 65 exclusions shown separately. The businesses were discovered through searches and public descriptions of independent or local trading, across retail, hospitality, trades and professional services in all four UK nations. This is a convenience cohort: employee counts were not independently verified, and the findings do not estimate the whole UK small-business population.
## What did we find in the homepage HTML?
The clearest finding is variation in basic page structure. The presence of structured data and a single main heading differed across the successfully measured pages. A question-led heading was less common, but that is not automatically a defect: a shop homepage can help customers without reading like a long question-and-answer guide.
| Feature | Pages | Base | Share |
|---|---:|---:|---:|
| At least one JSON-LD script block | 155 | 236 | 65.7% |
| Exactly one H1 in the HTML reader | 128 | 236 | 54.2% |
| At least one question-led heading | 67 | 236 | 28.4% |
| At least one malformed JSON-LD block | 2 | 236 | 0.8% |
| A detected noindex directive | 5 | 236 | 2.1% |
The structured-data observation covers JSON-LD scripts in the fetched HTML. It does not count every possible markup format, nor markup that JavaScript might add later. A block being present also does not establish that its claims are correct, complete or eligible for a search feature.
## How were the websites selected and fetched?
We selected search-discovered business websites only after recording a public source for their location and independent or local trading description. We excluded institutions, national chains, directories and unclear locations. The original research keeps the query, source, date and eligibility explanation. A single domain represents each business, with its www and non-www variants deduplicated.
The sample deliberately spans different services and locations, but the quotas are not population weights. In the coverage tables, total means attempts, scored means complete measured observations, and skipped means exclusions. Search discovery may favour businesses that already have a useful online presence. The words independent, family-run and local are descriptions from public sources, not audited employee or turnover classifications.
| Sector | Total | Scored | Skipped |
|---|---:|---:|---:|
| hospitality | 62 | 47 | 15 |
| professional services | 77 | 62 | 15 |
| retail | 85 | 65 | 20 |
| trades | 77 | 62 | 15 |
| Nation | Total | Scored | Skipped |
|---|---:|---:|---:|
| England | 152 | 127 | 25 |
| Northern Ireland | 30 | 22 | 8 |
| Scotland | 56 | 42 | 14 |
| Wales | 63 | 45 | 18 |
Collection used verified HTTPS, a named study crawler and public-address DNS checks. It read robots.txt before requesting a homepage or llms.txt, respected applicable access rules and crawl delays, and shared host skips across workers. Requests were spaced by at least one second globally and three seconds per host, with at most two domains in progress. It did not use private pages, cookies, forms, a proxy or security bypasses.
## What could not be measured?
The 65 exclusions are part of the result. They are not zero scores, evidence that a business is poor, or proof that the site is unavailable to its customers. A refusal, timeout or DNS failure describes this research request. The collector did not retry failed hosts through another route.
| Grouped exclusion reason | Attempts |
|---|---:|
| denied challenged rate limited or server error | 13 |
| robots policy unavailable or unusable | 7 |
| request timeout | 16 |
| robots refusal | 1 |
| homepage unavailable or non html | 1 |
| dns failure | 14 |
| tls or certificate failure | 11 |
| other fetch failure | 2 |
This is a complete-case dataset. If the optional llms.txt check failed after a successful homepage request, the full observation was excluded rather than silently mixing partial records into the score tables. That conservative choice can exclude otherwise reachable homepages. Our own crawler being refused does not demonstrate that a verified search-provider crawler is refused too.
The study also did not execute JavaScript, inspect every page of a website, check real customer questions in assistants, observe referrals, or measure sales. Those need separate experiments. A homepage result cannot stand in for a full business website or an answer about a particular product or service.
## Did websites block ChatGPT or other AI crawlers?
We evaluated the fetched robots policy against separate search and training tokens. This table is conditional on the successfully measured homepages. A zero means that no disallow rule matched within that measured set; it does not mean all attempted sites can be reached by the named service.
| Token | Disallowed | Base |
|---|---:|---:|
| OAI-SearchBot | 0 | 236 |
| GPTBot | 6 | 236 |
| Claude-SearchBot | 0 | 236 |
| ClaudeBot | 4 | 236 |
| PerplexityBot | 0 | 236 |
| Googlebot | 0 | 236 |
| Bingbot | 0 | 236 |
[OpenAI documents search and training as separate crawler roles](https://developers.openai.com/api/docs/bots). OAI-SearchBot relates to ChatGPT search; GPTBot relates to potential model-training use. A business can make different choices for those uses. This study did not impersonate those bots or test their published IP ranges against hosting firewalls, so it cannot establish provider-level reachability.
## What did llms.txt tell us?
Among the 236 measured observations, 73 returned a text response at llms.txt, and 69 passed the site's format validator. There were 161 absent responses, 2 responses classified as not a text file, and 0 checks not requested because robots rules denied that path. These states remain separate in the download.
The [llms.txt proposal](https://llmstxt.org/) describes a concise content guide for agents. Finding a file is not evidence that an assistant has used it or that a business receives citations. [Google's own AI-feature guidance](https://developers.google.com/search/docs/appearance/ai-features) says its established search practices remain relevant, without a special additional optimisation requirement. This study does not establish that adding llms.txt improves ranking, traffic or revenue.
## How should the readiness scores be interpreted?
The median schema component score was 54.0/100, with a range of 0 to 97. The median answer-readiness component was 46.0/100, ranging from 3 to 86. Each summary uses the same 236 measured pages.
These are AEO Lookup's editorial heuristics, not an externally validated visibility scale. The answer scorer rewards explanatory features such as question headings, dates, source links and sufficient text. That can favour an article over a short but useful booking or retail homepage. We therefore lead with observable component counts, not a league table or a pass/fail judgement about businesses.
The schema and answer functions match scoring source commit e2afbef; their exact file hashes are retained in the [method file](https://aeolookup.com/research/uk-business-2026/method.json). The collection omits the full API's firewall-style crawler probes, so these results are not the public audit API's overall score. They are not percentages of an assistant's knowledge or probabilities of being recommended.
Language is another limitation. The question-word heuristic is English-oriented and can undercount questions in Welsh or another language. Earlier collection rows did not retain the lang attribute, so unknown includes both unrecorded values and missing declarations. Location is never used to infer language.
| Recorded HTML language | Measured pages |
|---|---:|
| af | 1 |
| cy | 2 |
| en | 64 |
| en-GB | 68 |
| en-US | 17 |
| en-gb | 2 |
| unknown | 82 |
## What should a business owner check first?
Use the findings to prioritise checks on your own website, while keeping technical improvements separate from actual customer outcomes.
1. Check that important pages return usable text and that intended search access is not restricted by a mistake in robots, indexing directives or the hosting layer.
2. Make the business, service, location and useful customer information clear. A coherent main heading and accurate structured data help describe the page; avoid adding irrelevant schema just to chase a score.
3. Answer genuine customer questions on the pages where those answers belong. Do not turn every homepage into a long generic article.
4. Run a separate, repeatable question log in assistants and retain the returned source links. Track visits independently when suitable analytics are available.
5. Recheck after a meaningful change, keeping the same conditions where possible. Do not describe a score movement alone as new leads or market share.
The free audit and crawler checker can help inspect a particular site. Our guides on [whether ChatGPT can see your website](https://aeolookup.com/ai-for-small-firms/can-chatgpt-see-my-website/) and [measuring AI search visibility](https://aeolookup.com/ai-for-small-firms/measure-ai-search-visibility/) explain the distinction between a technical check and an observed citation.
## Can the findings be checked or reused?
Download the [pseudonymous observations CSV](https://aeolookup.com/research/uk-business-2026/observations.csv), [column and limitation README](https://aeolookup.com/research/uk-business-2026/README.txt) and [method fingerprints](https://aeolookup.com/research/uk-business-2026/method.json). Blank cells mean unavailable, not zero. The public rows omit names, domains, page bodies and contact details; consequently the public download cannot independently map a score back to a named business.
The editorial record retains searched eligibility evidence, response hashes, timings and the original scoring-source hashes. Before publication, a fixed systematic subset of 20 successfully measured domains was selected in alphabetical order. A second HTML parser and editorial review cross-checked H1 counts, JSON-LD block parsing and language extraction, retaining any changed responses or disagreements rather than replacing inconvenient examples.
Cite this report as a dated convenience-cohort study of homepage technical readiness by AEO Lookup. Preserve the denominator, exclusion count and scope when quoting a percentage. Do not relabel it as the proportion of all UK businesses visible to AI. A later measurement can extend the cohort or study assistant citations, but it should have a separate date and method rather than silently rewriting this snapshot.
## What are the common questions about this study?
### Does this study show whether ChatGPT recommends a business?
No. It measures selected homepage HTML features and declared robots policies. It does not test real assistant citations, recommendations, firewall reachability or referral visits.
### Is this a representative survey of UK small businesses?
No. It is a search-discovered convenience cohort of 301 attempted domains, selected from public descriptions of independent or local businesses. Company headcounts were not independently verified and no national weighting was applied.
### Were blocked or failed sites given a zero score?
No. The 65 excluded attempts are shown separately. Scores and homepage-feature percentages use only the 236 successfully measured pages; unavailable values remain blank in the download.
### Does a missing llms.txt file mean an AI visibility problem?
No. The file is a proposed content guide, not proof of search eligibility or a citation guarantee. This study records its presence and format without claiming a measured visibility benefit.
---
General information for UK small firms, not legal, tax or financial advice.
