Whether AEO and GEO work is paying off can be watched, one buyer question at a time. On 5 October 2026 we fixed a set of 39 questions a buyer might put to an AI assistant about what we sell. Here is what the set measures, how we read it, and how a citation is traced to a conversion.
What AEO and GEO change, and what they do not
Answer engine optimisation (AEO) is the work of making a page the answer an engine lifts for a question. Generative engine optimisation (GEO) is the work of being mentioned and cited when an assistant writes its own answer. Both are content, structure and authority work. Neither has anything to do with automated bidding in ad platforms.
The conversion question follows from that. An AI answer often settles the question without a click, so a conversion rate measured only on sessions misses the first step. We measure three things in order: is the brand present in the answer, is the right page cited, and does that page convert the visits it gets.
- 01PresenceIs the brand mentioned in the answer at all?
- 02The right page citedThe tagged page, not one with no route to a proposal
- 03ConversionWhat visits referred by AI assistants go on to do
The prompt set: 39 questions in five lanes
Each prompt is a question a real buyer could type, tagged with the lane it belongs to, the audience asking it, the market (the United States, for this set) and the one page on our site that should be cited if the answer is working.
| Lane | Prompts | Example question | Page we expect cited |
|---|---|---|---|
| White-label for agencies | 12 | Who can manage PPC campaigns white-label for my agency’s clients? | The matching white-label service page, or the agencies hub |
| Marketing services | 10 | What is a good SEO agency for a SaaS company? | The matching service page, such as SaaS SEO |
| Web builds | 6 | Who can migrate my store from Magento to Shopify? | The matching development page |
| Design | 3 | Who offers UI and UX design for a website or online store? | The matching design page |
| AI and the brand itself | 8 | How much does PixelCrayons charge? | The AI service page, or our company, trust, pricing or audience page |
Twelve prompts are asked as an agency and 27 as a brand. Thirty-three never mention us, so our appearance has to be earned. Six name PixelCrayons directly: what we are, whether we are legitimate, who owns us, what we charge, and two comparisons with a named competitor.
How we read each prompt
A reading is a record of what an assistant said on a given date, not a ranking. We built the set so that each prompt can be read the same way every month:
- Presence. Is PixelCrayons mentioned in the answer at all?
- Citation. Is a pixelcrayons.com page cited, and is it the page the prompt is tagged with, or a different one?
- Company. Which other domains are cited, and for which part of the answer: as a provider, as a source of advice, or as a directory?
- Accuracy. For the six branded prompts, is what the answer says about us correct: ownership, services and how pricing works?
- Change. Compare with the previous month, prompt by prompt, and note what changed on our side in between.
Two rules keep the readings honest.
Branded and unbranded prompts are never pooled into one figure, because a question that names us gives us an easier chance to appear. And a single reading proves nothing: answers vary from run to run, so a movement counts only when it holds across months. The full recording protocol is in our AI visibility audit post.
From citation to conversion
The page tag is what connects visibility to revenue. If an answer about white-label PPC cites a blog post instead of the service page, the reader lands somewhere with no route to a proposal. That is a page problem we can fix: strengthen the page that should be cited, and link the cited one to it.
On the conversion side, we look at visits referred by AI assistants in analytics, by landing page, and at what those visits go on to do. The aim is to compare pages, not to claim that AI visitors are worth more than other visitors in general. Our AI search optimisation work runs on this loop.
The clearest evidence we have published is a client programme rather than our own site: a UK B2B SaaS platform whose organic traffic grew 127% in five months, with citations in Google AI Overviews on commercial queries. The figure belongs to the whole programme. The record is in the SaaS case study.
Readings, and when this is the wrong measure
Readings are taken monthly and dated. This post sets out the method; it does not report a result, because a result belongs with its date and its prompt. When readings are added, each will say which month it covers.
A monitored prompt set is the wrong tool when:
- Your buyers do not use AI assistants to choose in your category. Ask your sales team before building a set.
- You sell locally. Map listings and reviews usually matter more than a national prompt set.
- You want a number by next month. Thirty-nine prompts are a sample, not a census, and the trend needs several readings.
