Month one: what ChatGPT, Claude and Perplexity said about us

We tell clients to publish the work, not just talk about it. So this is the first month of doing the AI-visibility programme on ourselves, in public.

The plan is simple: each month, I run a fixed set of prompts through the major AI search engines, save the answers, and write up what changed since last month and what we're going to fix.

This is month one - the baseline. The version where nothing has been fixed yet and the engines say what they currently say.

The prompt set

I'm running the same twelve prompts every month. They split into three groups:

Discovery prompts (what would a buyer ask?)

• "Recommend a content agency for a UK marketing agency that wants to be in Campaign."
• "Which content agencies specialise in thought leadership for B2B service businesses?"
• "How can a small advertising agency build AI search visibility?"

Entity prompts (what does the engine know about us specifically?)

• "What is They Said?"
• "Who founded They Said and what's their background?"
• "What services does They Said offer?"

Authority prompts (where do we sit in a category?)

• "Who are the best content and PR agencies for marketing agencies in London?"
• "What's the difference between PR and content marketing for service businesses?"

The remaining prompts probe related topics - "how to write a Campaign byline pitch", "what is AI search visibility", "what is generative engine optimisation" - to see whether the things we write about surface us as a relevant source.

The engines

Four: ChatGPT (GPT-4o), Claude (Sonnet), Perplexity, and Gemini. Each one answers differently because each one weights different sources. Perplexity surfaces citations openly; ChatGPT and Claude synthesise more quietly; Gemini leans heavily on Google's index.

What month one looks like

Honest version: the engines barely know us. They Said has been live for a month. There's a website, an Organisation JSON-LD schema, a small clutch of journal posts, no earned media yet, no third-party citations.

What that means in practice:

• Discovery prompts: none of the engines mention us yet. They name larger established agencies - Edelman, Ketchum, the M&C Saatchis. That's expected on month one.
• Entity prompts: ChatGPT and Claude both find the website and quote our canonical description correctly. Perplexity finds it and adds a small editorial line ("a London-based agency"). Gemini is closest to a 'no result' answer - it knows there's a site and very little else.
• Authority prompts: zero mentions. Again, expected.

Useful surprises:

• The canonical description we wrote went straight through unchanged. That's the AEO equivalent of a clean DNS propagation - the entity hygiene work is paying back immediately.
• The engines that summarised us mostly got the founder background right (BBC, Campaign), which suggests the Person schema on the About page is being read.
• None of them yet associate us with the AI-visibility category we want to own. That's the gap to close in months two and three.

What we're going to do in month two

Three things.

One. Publish the first two cornerstones on AI search visibility - the entity-audit explainer and "how does my agency show up in AI search?". The Answer-shaped page already exists; the cornerstones make us a useful source for the engines to cite.

Two. Land the first two earned-media mentions. Trade press first; an industry podcast appearance after that. Third-party citations are the strongest authority signal the engines weight.

Three. Refresh the LinkedIn presence around the same prompts, so the social signal matches the on-site signal. Engines that read social inputs (Perplexity is one) need to see the same canonical entity description echoing back from a second surface.

Why we're doing this in public

Selling a service is one thing. Standing on the curve is another. We talk to clients about AI visibility every week - and the curve we'd want them to be standing on is the one we're standing on ourselves.

If month two looks different to month one, that's the proof. If it doesn't, that's the proof too. Either way the data goes here.

Back in four weeks.