Most teams treat AI visibility as a yes or no question: did the model mention us? After four weeks of controlled runs, that framing looks like the wrong one. Being named third in a list of six is closer to being invisible than to being recommended, and the levers that move you up the list are not the ones that get you into it.
The setup #
We picked 40 prompts a buyer would plausibly ask before choosing a payment gateway, from broad discovery to narrow comparison. Each prompt ran daily in three markets, and we logged the full answer, the order brands appeared in, and every source cited. That produced roughly 900 answers per week for a single category.
Three variables changed across the runs:
- the phrasing of the prompt,
- the market it ran in,
- and whether a rival had published something new in the sources the model tends to cite.
Everything else stayed fixed.
| Metric | Result |
|---|---|
| Answers naming the same top three | 61% |
| Domains supplying half the citations | 12 |
| Click intent when named first | 2.4× |
Position beats presence #
Across every market, moving from fourth to second in a list changed downstream click intent more than appearing at all did. A brand named last in six is read as a hedge, not a recommendation. That is the single most useful reframe we took from this run: GEO is a ranking problem, not a mention problem.
The mechanism is unglamorous. Models assemble a shortlist from a handful of sources they trust for the category, then order it by how confidently those sources make a claim. Vague positioning gets you listed. Specific, verifiable claims get you listed first.
Where AEO does the work #
Answer engine optimisation is where most of the winnable ground sits. When we rewrote a comparison page so each claim sat in its own short paragraph with a number attached, the same page started being quoted directly inside answers within nine days. Nothing about the domain changed. Only the shape of the writing did.
If a sentence cannot be lifted out of the page and still be true on its own, an answer engine will not lift it.
A quick way to test a page #
Strip a page down to its claims and see whether each one survives on its own:
# Pull every sentence containing a number and read them cold.
grep -oE '[^.]*[0-9]+[^.]*\.' page.txt | sed 's/^ *//'
If a claim needs the paragraph around it to make sense, rewrite it. Inline markers like <strong> or a stray <em> will not save a sentence that has no standalone meaning.
What SEO still decides #
Classic SEO has not stopped mattering, it has changed job. The pages that rank in traditional search are disproportionately the pages generative engines cite, so organic rankings now act as an input to AI visibility rather than the end goal. Technical health, internal linking and topical depth remain the price of entry.
The practical split we now recommend:
- SEO to be findable.
- AEO to be quotable.
- GEO to be recommended.
They are sequential, and skipping the middle step is the most common mistake we see.
How to run this yourself #
Pick 30 to 40 prompts a buyer would genuinely ask, run them daily in the markets you sell to, and log position rather than presence. Change one thing at a time and give it two weeks. Most teams see their first order change well before they see a traffic change, which is exactly why tracking answers matters.
Methodology notes and the raw prompt list are available on request.