Why AI Answer Rankings Churn: A 52-Day Measurement

Most advice about AI visibility assumes the leaderboard holds still long enough to climb. We tested that assumption. On 21 July 2026 we ran 8 Dubai buyer questions through Google AI Mode and saved every answer. On 12 September we ran the identical 8. The field turned over completely, and nothing visible had changed about any of the agencies involved.

The short version

We measured the same questions twice, 52 days apart. Push Group fell from a 50% citation share to 0%. Digital Nexa fell from a 37.5% named share to 0%. AIIMS Group fell from 37.5% cited to 0%. Only 20% of the July field appeared anywhere in September, and the citations that replaced them scattered across one-off domains rather than concentrating on a new leader. No agency in that set changed its pages, its pricing or its positioning in the interval. The engine re-sampled and returned a different world.

What we measured

Agency	July citation share	September citation share	Change
Push Group	50%	0%	Total loss
AIIMS Group	37.5%	0%	Total loss
Houses of Growth	37.5%	0%	Total loss
ThatWare	37.5%	12.5%	Reduced
Bright Ideas	25%	0%	Total loss
Lengreo	0%	25%	New entrant
GCC Technologies	0%	12.5%	New entrant
Ivanooo	0%	0%	Unchanged

Four reasons a leaderboard moves without you

The index is re-crawled, not frozen

Generated answers are assembled at query time from whatever the engine currently holds. A source that was fresh in July can be outranked in September by a page published in August, with no penalty applied to the first.

Retrieval is probabilistic, not ranked

An AI answer is not a top-10 list with stable positions. The engine samples sources that fit the question, and small changes in phrasing, locale or model version shift which sources clear the bar. The Princeton GEO research, Aggarwal et al., showed that passage-level changes alone moved visibility by up to 40% under controlled conditions, which tells you how sensitive the selection is.

The category itself is young

Dubai AEO is a market where new service pages appear weekly. In September our panel surfaced Plus Point Digital, Gateway Marketing Digital, Decipher Agency and SEO Tech Experts, each cited once. A thin, contested category re-sorts faster than an established one.

Concentration decays into a long tail

July concentrated citations on a handful of agency estates. September spread them thin. That pattern matters more than any single position: when the engine stops maintaining a shortlist, no one on it can claim a durable seat.

What a buyer should do about it

1. Stop buying positions and start buying coverage. A position is a reading on a date. Coverage is the asset that keeps getting re-sampled into new answers.
2. Set the review cadence at 60 days, not annually. Anything slower and you cannot tell a real gain from a re-sample.
3. Demand two dated readings from any vendor before signing. One reading proves nothing about method.
4. Treat a sudden gain with the same suspicion as a sudden loss. Both can happen without any work being done.
5. Measure the questions, not the rank. Which of your buyer questions return you at all is a more stable signal than where you sit inside one answer.

What churn does not mean

It does not mean the work is futile. Coverage held over time is exactly what survives re-sampling, because a site that answers a subject completely keeps qualifying for new answers as the engine re-draws them. Topic Authority is the durable layer; position is the volatile one. What churn kills is the guarantee, not the discipline. The limits worth knowing before you buy set out the rest.

Questions buyers ask

Does this happen on every engine? We measured Google AI Mode. The mechanism, query-time assembly from a changing index, applies across generative engines, but the counts do not transfer. Measure the engines your buyers use.

Could the churn be caused by a model update? Possibly, and we cannot see inside the engine to confirm it. What we can say is that the change was wholesale rather than gradual, which is more consistent with a re-sample than with gradual page decay.

If I am winning today, how worried should I be? Worried enough to have a second reading. Push Group held the strongest position in our July panel and scored zero 52 days later.

Does publishing at a higher cadence protect me? Freshness helps retrieval, and our own AI Search Visibility work treats it as one lever among several. It does not lock a position, because nothing does.

Why did Ivanooo stay at 0% across both readings? Because we publish our own numbers, not only our clients'. As Firoz Azees puts it: "a measured zero is a work order; a flattering estimate is a story you tell yourself." Our zero is stable for the boring reason: we have not yet built the coverage these questions reward.

How many readings make a trend? Two tell you whether something moved. Three or more tell you whether the movement has direction. We now hold three dates on this panel and the only stable line in it is our own zero.

What should I ask an agency about churn? Ask what their clients' numbers did over the last 60 days, including the ones that fell. An agency that has never seen a client decline has either not measured twice or is not showing you both readings. How to verify those claims covers the rest of the diligence.

At Ivanooo, Firoz Azees runs Distinctiveness Engineering for the AI-answer era: measuring who the engines name, cite and recommend, then engineering the gap between listed and chosen. Start with a free AI visibility check and take your own first reading.