GEO Isn't a One-Off: Why Your AI Ranking Drops

Three months into GEO (Generative Engine Optimization), many owners ask the same question: we used to show up in Doubao and DeepSeek, and now we're gone — was it all for nothing?

The conclusion first: it isn't wasted effort — it's a misunderstanding of GEO. An AI answer isn't a leaderboard that stays put once ranked; it is regenerated on every query. So GEO was never a "do it once" project; it's something that needs ongoing upkeep.

One boundary though: this doesn't mean staring at it daily or publishing every day. Rhythm matters more than intensity — get the frequency right and the workload is small; get it wrong and no amount of effort helps.

First, an unsettling experiment

In 2026, SparkToro and Gumshoe ran an experiment: 12 prompts × 3 platforms × 600 volunteers, 2,961 runs in total. The finding: for the same question, seeing the exact same list of brands again is almost impossible. (The experiment is reported via media.)

Scope note: this experiment measured the stability of answers, not whether AI is reliable; it also found that brands appearing repeatedly still reflect a real pattern. So "it's different every time" doesn't mean "there's no pattern".

Use that finding in two halves.

• A single screenshot means nothing. Not seeing yourself in AI one day doesn't prove you're out; seeing yourself on another day doesn't prove you're safe either.
• Only the trend means something. If you're absent across several platforms for weeks on end, that's a real problem.

Most people do it backwards — treating one screenshot as a conclusion, then either celebrating or panicking for nothing.

Why is it different every time?

Because AI's answers are computed on the spot, not pulled from a fixed leaderboard.

Google's AI Mode, for instance, is documented to perform query fan-out: it splits your question into up to 16 sub-queries, retrieves each separately, then assembles one answer.

The split isn't fixed, the retrieved content changes, timestamps change — so the output inevitably fluctuates. That's the mechanism, not a flaw in your content.

Once you understand this, a lot of anxiety dissolves: you aren't competing against a fixed leaderboard, but against a candidate pool that is recombined every time.

So the right rhythm for GEO is three things

Not "publish constantly" — three actions at different frequencies.

• Monitor — fixed questions, fixed frequency. List the 10 to 20 questions customers are most likely to ask, run them through your target AI engines weekly or fortnightly, and log whether you appear, where, and who gets cited. Watch the trend line, not a single point.
• Update — content isn't finished when it's published. "What Gets Cited", accepted at SIGIR 2026, found across 252,000 controlled experiments that a fresher timestamp has a consistent positive effect. That means stale content gets squeezed out over time, and core pages need a plan for updates.
• Expand — grow outward from the core questions. Nail the most central dozen or so first, then move to adjacent questions. Doing it in reverse (blanketing everything from day one) usually leaves every point shallow.

Scope note: the conclusion in "What Gets Cited" comes from controlled experiments; its percentages represent AI's relative preference among candidate content, not a promise of real-world results. But it does show at least that freshness is a variable that gets counted.

What to look at in monitoring — and what not to

This section matters, because many people are misled by the tools themselves.

Bing's official AI Performance dashboard, launched in 2026, states three limits in its own documentation.

• Total citations do not include position or presentation within an answer;
• The daily average of cited pages does not represent ranking, authority, or any single page's role in one answer;
• No click data is provided.

The official framing is "a first step toward GEO tooling" — in other words, these are observational metrics, not a scoreboard. The same applies to Citation Share in that dashboard: the documentation explicitly calls it observational, not a ranking system or a competitive scoreboard.

One more detail is easy to misread: the Grounding Queries in that dashboard are not the user's own words — they're retrieval phrases generated internally by the search engine. Treating them as "this is how users asked" leads to wrong conclusions.

The takeaway: tools are fine to look at, but read them within the limits they declare. Treating observational data as a performance review is the most common misuse in this field.

Three common questions

1. How often should we run monitoring?

It depends on how long your question list is. A general suggestion: core questions weekly or fortnightly, the full list monthly. The key word is fixed — same questions, same frequency, same recording method, so the data can be compared. If you switch questions every week, you get a pile of fragments that never forms a trend.

2. How often should content be updated?

No single answer, but there's a test: update when the facts on the page change. Product specs change, industry data updates, policy shifts, company details change — all of those need to be synced. You don't need to change dates pointlessly just to look fresh — that achieves nothing and can make your content messier.

3. If we switch providers or take it in-house, can we take the data with us?

This is something to settle before signing: the question list, historical monitoring records and content update logs should belong to the company itself. If the provider locks that data in its own dashboard with no export, then the day the engagement ends, your GEO work restarts from zero. Better to write this into the agreement up front.

In closing

To sum up: GEO isn't a project, it's a habit.

Projects have delivery dates; habits don't. Building your website and content is just preparing the ground; monitoring on schedule, updating as needed and expanding step by step is what actually keeps it working.

More practically: the hard part isn't the first pass, it's building a lightweight process you can keep running. Two hours a week sustained for six months beats a two-week "big project" that then sits idle.

You can do one thing right now: write down the 10 questions your customers are most likely to ask, as your first monitoring list. Run the same list again next month and you'll have your first set of trend data.

Do you have a fixed monitoring list, or do you only ask AI when it crosses your mind? Tell me in the comments.

Yihang Network, based in Shenzhen, previously worked in cross-border B2B and Alibaba.com-related services, and expanded into GEO from 2024. Our view: the value of GEO isn't in any single optimisation, but in whether it can keep running.

Data and scope note: The SparkToro and Gumshoe experiment (2026) is reported via media and is a third-party finding; the conclusion in the SIGIR 2026-accepted paper "What Gets Cited" comes from 252,000 controlled experiments and reflects AI's relative preference among candidate content under controlled conditions; the characterisation of the Bing AI Performance dashboard comes from its own official documentation. None of this represents real-market outcomes or any single company's results, and none of it constitutes a promise of timelines or business results. The methods described raise the probability of being retrieved and cited; they do not guarantee a company will appear in AI answers.

1. SparkToro × Gumshoe, AI answer stability experiment (2026, reported via media)
2. SIGIR 2026 (accepted), "What Gets Cited", a controlled-experiment study
3. Google Search Central, official documentation on AI Mode and query fan-out
4. Bing Webmaster Tools, official documentation for the AI Performance dashboard (2026)