We asked AI models for a Suffolk web agency 400 times. These are the firms they named.

Between 3 August and 7 September 2026 we put the same 44 questions to two AI models once a week and recorded which web design firms each answer named. Forty of those questions never mentioned us. Across the five runs that is 400 unbranded calls, and six Suffolk firms turned up in them: one was named 65 times, then 45, 34, 33 and 24. Ours was named 6 times.

We came last. That is the finding, and the reason the rest of this is worth reading is that we have the run-by-run data behind it rather than an opinion about what AI search rewards.

The other five are real Suffolk web design and digital firms, tracked by name in our own data and left unnamed here. Publishing a table of our competitors' visibility scores is not the point of the exercise, and their numbers do not need our commentary attached to them. The shape is what matters, and the shape is not flattering to us.

How the test was run

Forty-four prompts, put to OpenAI and to Google Gemini with web search switched on, once a week for five weeks. Four of the prompts name Signal Red Studio directly. The other forty do not, and they are the ones that count, because a model naming a business it was told the name of proves nothing.

The prompts fall into four groups. Local discovery asks for a supplier in a place: I need a web designer in Stowmarket, Suffolk, who should I contact? Service asks who does a particular job in the region. Buying asks the questions someone asks before they commission anything, like how much a website should cost or whether to use an agency or a freelancer. Informational asks for an explanation, and each one ends with an instruction to cite the sources behind the answer.

Every answer is parsed for two things that are not the same. A domain is consulted when the retrieval layer fetched or considered it, and cited when the model actually attached it to a sentence in the answer. Cited is the harder number and it is the one used throughout this piece.

There was no run on 31 August. The scheduled job failed on an outbound network error, so the series has a five-week span with four weekly gaps and one fortnightly one.

Almost all of it happens on one kind of question

Split the 400 unbranded calls by what was asked and the pattern is not subtle.

• Local discovery, 130 calls. The six firms picked up 153 citations between them, because a single answer usually names several.
• Service, 90 calls. 49 citations.
• Informational, 120 calls. 5 citations, every one of them to the same firm.
• Buying, 60 calls. Zero. Not one local firm, from any of the six, in five weeks.

The informational number is the one that should worry anybody running a content programme. Those twelve prompts ask about local SEO checklists, Google Business Profiles, Core Web Vitals, schema markup, how long SEO takes, headless CMSs, React against WordPress, accessibility obligations and GDPR. We have published an article on nearly every one of them. Across 120 chances to be quoted as a source, the models cited our work zero times, and cited the six firms combined five times.

Generic advice content is the standard local SEO play, and on this evidence it is close to worthless as a route to being named by a model. The question that produces a recommendation is the one that asks for a supplier in a place.

Buying questions returning nothing at all is the sharper version of the same point. When somebody asks a model what a website should cost or how to choose an agency, it answers with general guidance and national sources. It does not hand over a shortlist of local firms, even when the prompt asks for credible options by name.

Four runs at zero

Our own six citations all landed in the final run, on 7 September. The first four runs, 320 unbranded calls, returned nothing at all.

Of the six, four came on local discovery questions about Stowmarket and Bury St Edmunds and two on service questions about custom software in Suffolk and local SEO around Ipswich. All six came from OpenAI. Gemini did not name us once on an unbranded prompt in five weeks, while it named one of the other firms 25 times.

One run is a sample, not a result. Model answers are non-deterministic, the same prompt returns different sources on the same day, and a jump from zero to six could reverse next week. What makes it worth reporting rather than sitting on is that four consecutive zeroes is a stable enough baseline that a break in it is visible.

What this does not measure

Querying a provider API with web search on is not the same as being a person using ChatGPT. The API has different retrieval backends, different system prompts, no user memory or personalisation, and none of the licensed data layers that exist only in the consumer products. ChatGPT's local answers draw on Foursquare Places and licensed Yelp data, neither of which is reachable through the API. For local questions in particular, this method measures something adjacent to what a real user sees rather than the thing itself.

So it is a relative, longitudinal signal. Six firms, the same prompts, the same models, week after week. It supports statements like "the most-cited firm in our set is named ten times more often than we are on the questions that matter" and does not support any claim about how many people saw either of us.

Anyone quoting a single week of this kind of data at you, their own or ours, is overreaching.

The name collision nobody plans for

Four of the 44 prompts name Signal Red Studio, and one of those asks about our CRM product. On five of the ten times that prompt ran, the model answered with an American security testing company that has a similar name and nothing to do with us.

Traditional search does not make that mistake, because a link either goes to your site or it does not. A model composing prose can name the right business and attach the wrong domain, and the person reading it has no way to tell. If your trading name is close to somebody else's, that is a specific and checkable risk, and the only way to find out is to ask the models and read what comes back.

The one number that is not inferred

Everything above is measured by asking models questions. The one source that is not is our own server log, which records what answer engines actually fetched.

Over the eight days to 7 September, agents that fetch a page in order to answer a live question hit the site 26 times: 22 from ChatGPT-User and 4 from Claude-User. Every one of them landed on the homepage. Separately, index crawlers hit the site 213 times and training crawlers 113 times, and neither of those tells you anything about whether you were cited.

Twenty-six is a small number and it is a real one. Each hit is a person who asked something and an engine that went and read our homepage to answer them.

What we changed

Three things, on the evidence above rather than ahead of it.

Town pages carry the specifics a local discovery answer needs, which is a named place, a real address, a phone number and a plain statement of what is done there. That work predates this data and the four Stowmarket and Bury St Edmunds citations are the first sign of it landing.

Pages are prerendered rather than rendered in the browser. AI crawlers do not run JavaScript, so a page that only exists after hydration is invisible to them regardless of how it looks to you. We found one of our own pages sitting at 1,226 characters of crawler-readable text this month and fixed it.

The measurement runs weekly and is kept whether it flatters us or not. This article exists because the honest chart puts us at the bottom of it, and a version that put us at the top would have had to be built from the branded prompts, which measure nothing.