How to Measure AI Search Optimization: Stop Watching Keyword Rankings

Whether AI search optimisation (Generative Engine Optimization) is working cannot be judged by keyword rankings — because AI doesn't give rankings at all. To judge results, look at three things: whether AI mentions you, where you sit when it does, and whether you or a competitor gets recommended first.

But those three metrics have limits too: they measure the probability of being cited, not closed business. Treating probability as a promise is the most common misreading in this field right now.

Why rankings stopped working as a metric in the AI era

Traditional SEO had a clear scorecard: which page your keyword ranks on, how much organic traffic grew, what the click-through rate is. In AI search, almost none of that applies — because AI doesn't give you a position, it gives you a paragraph. You may be written into it, or you may be entirely absent.

Pew Research tracked user behaviour in July 2025: when an AI summary appears at the top of the results page, only 8% of users click through to traditional search results; without that summary, the figure is 15%.

One scope note here: that 8% is a share of click behaviour, not of conversions, and it cannot be reversed into "doing AI search optimisation will get you the corresponding traffic". It says one thing only — users increasingly read the answer on the results page and no longer click through.

Flip it around: if users stop clicking, the meaning of "what rank am I" is heavily diluted. That's exactly where many companies get confused — the ranking is still there, so why have enquiries dropped?

Three metrics to watch

Swap ranking thinking for citation thinking, and three metrics matter:

Metric one: mention rate — does AI mention you at all?

In your own industry, ask AI the sentences your buyers actually use and check whether your brand appears. A LED lighting supplier, for instance, asks "which LED lighting suppliers in Shenzhen are reliable". Ask 5 to 10 questions and count how many answers mention you.

Metric two: citation position — where you sit when you are mentioned

Being mentioned isn't enough; position matters. AI usually highlights only a handful of names, and the probability of a buyer seeing the first versus the last is completely different.

Metric three: competitor share — who you appear alongside

More important is the side-by-side comparison: when AI recommends this category, which companies keep reappearing? If a competitor is in every answer and you're in none, AI has already formed a default answer for that category.

A counter-intuitive finding: optimisation helps the fifth-ranked page and hurts the first

This is the one I think every owner should know about — and it comes from the academic origin of the term GEO itself.

A KDD 2024 paper (Aggarwal et al., Princeton and IIT Delhi) tested this with 10,000 real queries across 25 domains. The result was unexpected: the same content optimisation produces opposite effects depending on where the page originally ranked.

Scope note: these are relative changes under laboratory conditions and a fixed query set, not a guarantee of real-market results. But it indicates a direction: in AI answers, an existing positional advantage gets weakened, and content with the evidence filled in has a chance to overtake.

The same paper also found that keyword stuffing is ineffective or even harmful in AI answers. That is the academic confirmation of where the division of labour between GEO and SEO begins.

Why I can't give you an exact number

Before writing this I checked the available data. The conclusion first: there is currently no authoritative statistic in this field.

One example: on "the share of B2B buyers using AI", for the same year I found six versions — 29%, 45%, 51%, 73%, 84% and 94%. More remarkable still, the 94% figure is attributed to Forrester in some articles and to 6sense in others — two different sources claiming the same number.

"The size of the domestic GEO market" is the same: from ¥250 million to ¥28.6 billion — a difference of two orders of magnitude.

So this article cites only two kinds of data: figures published by official bodies (such as CAICT) and academic research with clear provenance (such as the KDD paper). If someone gives you an exact "AI citation share", ask where the data comes from first.

Three questions about AI search optimisation results

How soon will we see results?

No timeline and no results are promised. What can be said clearly is the pattern: AI has to crawl, then understand, then accumulate trust. The nature of it is saving up, not buying ads — stop paying for ads and it stops; what's been saved stays.

Do we need to buy a monitoring tool?

Not at the start. The crudest method works: fix 5 to 10 questions, ask them once a week in Doubao, DeepSeek and ChatGPT, screenshot the answers, and compare month by month. That's a zero-cost baseline.

Could results actually go down after doing the work?

Possibly, but it's usually for another reason — a model update, a competitor investing more, or contradictions appearing in your own content. That's exactly why monitoring has to be continuous: a single snapshot doesn't show a trend.

In closing

To sum up: AI search optimisation has no leaderboard, only a probability of being cited. So the way you measure it has to change too — mention rate, citation position, competitor comparison, not what rank your keyword sits at.

One thing to settle in advance: these metrics measure probability, not a promise. Any service claiming it will "guarantee you appear first in AI answers" deserves a question mark before you sign.

Have you searched for your own product in AI? Did you appear, and where? Try it with one question and note the answer down — that's your baseline.

Yihang Network, based in Shenzhen, previously worked in cross-border B2B and Alibaba.com-related services, and expanded into GEO from 2024. Our approach starts with diagnosis — establishing where the brand actually stands across Doubao, DeepSeek, Qwen and ChatGPT, and only then discussing optimisation.

Data and scope note: The Pew Research figures cited here come from an observational study of click behaviour and do not establish causation; the KDD 2024 paper is a peer-reviewed conference paper and its percentages are relative changes under laboratory conditions and a fixed query set, not representative of real-market outcomes or any single company's results; the CAICT figures come from public white papers and coverage in authoritative media. The methods described raise the probability of being retrieved and cited; they do not guarantee a company will appear in AI answers, and no timeline or business result is promised.

References

1. Aggarwal et al., "GEO: Generative Engine Optimization", KDD 2024 (DOI: 10.1145/3637528.3671900)
2. Pew Research Center, study on AI summaries and search click behaviour, July 2025
3. CAICT, "2025 Generative Engine Optimization Industry White Paper", cited via People's Daily Online coverage, March 2026
4. CNNIC, "Report on the Development of Generative AI Applications (2025)", October 2025