Keyper: Cited in 90% of Tracked AI Answers

Keyper is a Dubai proptech platform competing for attention in one of the most contested search markets anywhere. The property portals in that market have spent a decade and a great deal of money owning the results page.

So we stopped competing only on the results page. This is what tracking and optimizing for AI answers produced across ChatGPT, Perplexity and Gemini.

The problem

In Dubai property, the classic results page is effectively spoken for. The portals hold the head terms, and a challenger buying its way in pays portal-level prices for portal-level competition.

But the questions that precede a property decision are not head terms. They are things like how a tenancy contract works, what Ejari registration involves, or whether rent can be paid on a credit card. Those questions increasingly get answered by an assistant rather than a list of links, and the assistant picks its sources on different criteria than a ranking algorithm does.

That is a genuinely different competition, and at the time we started it was one almost nobody in the market was measuring.

What we did

Built a fixed measurement panel first. 73 questions a real buyer or tenant would ask, run across ChatGPT, Perplexity and Gemini on a schedule, with every citation recorded by source URL. Fixed, because a panel that changes between runs measures nothing. This came before any optimization work, so there was a baseline to argue with later.

Wrote for extraction, not for length. Each target page answers one question completely near the top, in language a buyer would use rather than industry vocabulary. The test we applied was whether a paragraph still makes sense lifted out of the page entirely, because that is exactly what an answer engine does to it.

Fixed the entity layer. Consistent naming, structured data that matches what is actually visible on the page, and a canonical set of facts about the company that does not contradict itself across the site. Engines resolve who you are before they decide whether to quote you.

Retired thin inventory instead of deleting it. Around 1,590 listing pages were drawing impressions at roughly a fifth of one percent click-through, which is a lot of crawl budget for very little. Rather than removing them, we redirected them into a smaller set of genuinely useful area guides. No 404s on indexed URLs, so the accumulated equity moved rather than evaporating.

What happened

The brand is now mentioned in 90% of the tracked question set, with 271 citation events in a single month spread across 69 distinct URLs. That spread matters more than the total: citations concentrated on one page is luck, citations across 69 pages is a pattern.

The distribution by engine was lopsided. ChatGPT produced 86% of the citations, which is a useful planning fact and an uncomfortable one, because it means the visibility rests heavily on a single provider's retrieval behavior.

The pages earning citations were not the ones we would have guessed. Practical, transactional explainers, rent payment by credit card, tenancy contracts, a single neighbourhood guide, did the work. The corporate pages did not.

Referral traffic followed from seven different AI platforms in one month. The volumes are small in absolute terms, and we tell clients that plainly: this channel currently sends fewer visitors than classic search, and its value today is presence in the answer rather than clicks from it.

What we would do differently

Set the panel wider at the start. 73 prompts was enough to prove the pattern and slightly too narrow to segment by intent afterwards. We would open with more coverage and accept a noisier first reading.

Instrument AI referrals before the content work, not after. Knowing which assistant sent a visitor, and to which page, is worth more than the visibility score, and it is much harder to reconstruct retroactively.

Treat the engine concentration as a risk from day one. With 86% of citations from one provider, a change in that provider's retrieval behavior is a single point of failure. The mitigation is deliberately building coverage on the engines where you are weaker, which is slower and less satisfying than compounding where you already win.

Where this goes next

The transferable part is the sequence rather than the tactics. Measure before you optimize, fix the entity layer before the content, write answers that survive being lifted out of the page, and report the trend from a fixed panel instead of a score that flatters. It works in any category where buyers ask questions before they choose, which is most of them.