The Verdict: Every Rank-Matched AI Overview Citation Was Top 10, None From 11-20. Our Other Two Predictions Were Simply Wrong.

Friday verdict — closing this week’s GEO Lab arc. Monday we found that a widely-repeated “97% of AI Overview citations come from top-20 pages” stat isn’t in the study it’s credited to. Tuesday we pre-registered three predictions and a 24-query protocol to test the question directly. Collection ran long enough that this post covers both the raw numbers and the scoring in one pass, rather than splitting results and verdict into separate days — full data is below, not just the scorecard.

What’s the verdict on each of the three locked predictions?

Three predictions went out Tuesday with pass/fail lines written before a single query ran. Here’s how each one actually scored against 24 queries (12 topics × keyword/question phrasing, 4 intents) run against the classic AI Overview box only — not AI Mode, not ChatGPT, not Perplexity.

Prediction	Verdict	What the data showed
P1 — magnitude: top-10 share lands 55-70%+, top-3 cited ≥2x more than 11-20	✅ Confirmed, exceeded	Among the 87 citations matched to a tracked organic rank, 100% were top-10 and 0 fell in positions 11-20 — a ratio more extreme than “2x,” because the denominator is zero. But that 100% is of matched citations only; 52 of 139 total citations (37.4%) weren’t findable in the organic top 20 at all. Counting those as “not top-10,” the honest top-10 share of everything cited is 62.6% — inside our predicted band, for a different reason than we expected.
P2 — intent: Local triggers strictly least often; its rank pattern is weakest if it triggers at all	❌ Rejected	Local triggered an AI Overview in 6 of 6 queries — tied with Product and Informational, and ahead of How-to (5 of 6). Our pre-registered reject condition (“Local triggers as often or more often than another bucket”) is met exactly. The second half of the hypothesis — that Local’s rank pattern would be weakest — turned out true, but for a different reason than “weaker correlation”: see below.
P3 — phrasing: questions trigger an Overview more often than keywords; top-10 share gap between them stays under 10 points	❌ Rejected, reversed	Keyword phrasings triggered an Overview in 12 of 12 cases; question phrasings in 11 of 12 (the one miss: a how-to query about debt consolidation steps). Keywords out-triggered questions, not the other way around. The top-10 share gap between phrasing groups (67.1% keyword vs. 57.6% question) stayed under 10 points, which would have passed — but the trigger-rate direction already fails the prediction on its own.

Net score: one prediction confirmed in a more extreme form than forecast, two rejected — one of those in the opposite direction from what we bet. That’s a worse hit rate than last week’s arc, and it’s exactly the kind of result a pre-registered test is supposed to produce when your intuition about intent and phrasing turns out wrong.

Why did the magnitude prediction land more extreme than 55-70%?

We predicted a softer version of the debunked 97% figure — 55-70% — reasoning from the 5.6x gap we’d already measured on a different surface (Google AI Mode’s 46.2% top-3 retention vs. 8.3% sub-top-20 retention). What we found on the classic AI Overview box is stranger: the concentration isn’t soft at all once you can see rank, it’s absolute. Of 87 citations we could match to a position in that query’s own organic top 20, every single one sat in positions 1-10. Not one came from 11-20. The debunked Monday stat claimed “97% from top-20, 50%+ from position 1” with no study behind it — our directly measured number for position 1-10 among rank-matched citations is 100%, which is in the neighborhood of that fabricated figure, just measured honestly this time and specifically for 1-10 rather than 1-20.

The catch is the other 37.4% of citations: pages the AI Overview cited that never showed up anywhere in that same query’s organic top 20, at least not in a single-location, single-account snapshot. Those break down by intent in a way that tells its own story:

Intent	Total citations	Rank-matched (in organic top 20)	Top-10 share of all citations
Product	38	27 (71.1%)	71.1%
Informational	52	35 (67.3%)	67.3%
How-to	36	23 (63.9%)	63.9%
Local	13	2 (15.4%)	15.4%

Product, Informational, and How-to cluster in a tight 64-71% band — comfortably inside our predicted 55-70% range once you stop treating “matched” and “total” as the same denominator. Local is the outlier, and it’s the outlier that breaks P2.

Why did Local trigger just as often as everything else, and why did its citations barely rank-match?

We bet Local-intent queries (“near me” service searches) would trigger a classic AI Overview the least often, on the logic that Google tends to answer those with the Local Pack or Maps results instead. That logic was wrong: all 6 Local queries triggered an Overview, same as Product and Informational, and more than How-to’s 5 of 6. Local didn’t get replaced by the Local Pack — it got an AI Overview box in addition to whatever local results also rendered.

What we didn’t predict is what that box cited. For “which dentist near me is accepting new patients,” the Overview returned zero external links — just a text answer naming three dental clinics by name and address, each annotated “Google” as the source, with no crawlable URL at all. For “emergency plumber near me” (keyword phrasing), the box cited exactly one domain, a Korean business-directory site that doesn’t appear anywhere in that query’s organic top 20. Across all 6 Local queries, 11 of 13 total citations were either sourced to Google’s own business-listing data with no external URL, or cited an external domain that our organic snapshot never surfaced. The 2 citations that did rank-match were both position 1, for the same AC-repair query under both phrasings.

That’s a different finding than “Local’s rank-citation correlation is weaker.” It’s that for local-service queries, the classic AI Overview box appears to draw substantially from Google’s own Business Profile / Maps-adjacent data rather than the open web results it shows you on the same page — so asking “does it cite by rank” is close to the wrong question for this intent bucket. The organic top-20 and the Overview’s citation pool aren’t competing for the same slots; they’re drawing from different inventories. We pre-registered a test that assumed they were the same inventory, and the reject condition caught that assumption as wrong before we got to the more interesting mechanism underneath it.

Why did short keyword phrasings out-trigger full questions, the opposite of what we bet?

We predicted questions would trigger an Overview more often, reasoning from Monday’s own fact-check — the Semrush study we debunked for the 97% figure did genuinely find that AI Overview-triggering queries tend to be longer and more complex than average. “Longer and more complex” turned out not to mean “phrased as a question” in our 12-topic sample. Every one of the 12 keyword-phrased queries triggered an Overview; only 11 of 12 question-phrased ones did. The single miss was “how do I consolidate debt without hurting my credit score,” a how-to query that returned a plain organic results page with no AI Overview box at all, while its keyword-phrased sibling (“debt consolidation steps without credit impact”) did trigger one.

With only one miss out of 24 total queries, this result rests on a single data point and shouldn’t be read as “keyword phrasing reliably beats question phrasing” — it’s evidence against our specific directional bet, not evidence for a strong opposite effect. The second half of P3, the top-10 share gap between phrasing groups, actually held: 67.1% (keyword) vs. 57.6% (question) is a 9.5-point gap, under our 10-point bar. Phrasing didn’t meaningfully change how citations track rank once an Overview appears — it just didn’t move whether one appears in the direction we guessed.

So does the classic AI Overview box cite by rank?

For Product, Informational, and How-to intent: yes, about as cleanly as a rank-citation relationship can look. Every citation we could rank-match sat in the top 10, never in 11-20, across 85 matched citations in those three buckets. That’s a sharper pattern than the softened 55-70% we predicted, and in the same direction as (though not the same claim as) the fabricated 97% figure Monday’s fact-check killed.

For Local intent: the question doesn’t really apply. A third of this week’s citations overall, and 85% of Local’s citations specifically, pointed at something other than a page in that query’s organic top 20 — either Google’s own business-listing data with no external URL, or a domain our snapshot never saw ranked. Rank-citation correlation assumes both signals are drawing from the same pool. For local-service queries, they often aren’t.

One honest limitation carried over from Tuesday’s pre-registration: this is a single-location (Seoul, KR), single-account snapshot, one round, 24 queries. AI Overview rendering is known to vary by account history, location, and device, and we didn’t control for that here — the same caveat as every prior GeoParrot arc. A second round or a second location could move the exact percentages. What’s unlikely to move is the structural finding: when this box cites a page with a traceable organic rank, that rank is concentrated at the top; a meaningful share of what it cites has no traceable organic rank at all.

What should you actually do with this, surface by surface?

Classic Google AI Overview — Product, Informational, How-to queries

• Top-10 organic rank is still the lever to pull. We found zero citations from positions 11-20 in this sample. If you’re outside the top 10 for a query that triggers a classic AI Overview, our data gives you no path into that box through rank alone — you’d need the content to surface a different way (forum mention, video result, etc.) that we didn’t track here.
• Don’t assume top-20 is “close enough.” The debunked 97%-from-top-20 stat implied a wide safety margin. Our directly measured number suggests the real margin, where we could see rank at all, was top-10 or nothing — a much tighter bar.
• Expect roughly a third of citations to be untraceable to rank. Don’t build a GEO strategy that assumes every AI Overview citation came from beating a rank threshold; plan for a meaningful share to come from elsewhere (a different surface, a format our organic scrape doesn’t capture, or simple snapshot variance).

Local-service queries

• Don’t optimize local-service pages for classic AI Overview rank-citation the way you would Product or How-to content. In 11 of 13 citations we logged, the Overview wasn’t citing anything that showed up in the organic top 20 at all.
• Treat Google Business Profile data as the more likely input. One Local query’s Overview named three businesses by name and address with “Google” as the only cited source, no external link in sight — consistent with Maps/GBP data, not web-indexed content.
• An AI Overview appearing on a “near me” query doesn’t mean the Local Pack disappeared. All 6 of our Local queries triggered an Overview; this isn’t an either/or with local-pack results.

Reading any “AI Overview cites by rank” claim from here on

• Ask what share of cited URLs the study could actually rank-match. A headline percentage that silently excludes unmatched citations (ours would have read “100% top-10” instead of the honest 62.6%) overstates the pattern.
• Ask whether the sample includes local-intent queries, and whether they’re reported separately. Blending Local into an aggregate hides the mechanism difference we found here.
• Treat phrasing-based intuitions (questions trigger more than keywords, or vice versa) as untested until measured. We had a specific, sourced reason to expect one direction and still called it backwards on a 12-vs-12 sample.

What’s next week’s question?

This week’s cleanest open thread is the Local-intent mechanism: if the classic AI Overview box is drawing from Google’s own Business Profile / Maps data for “near me” queries rather than the organic web index, does optimizing a Google Business Profile listing move AI Overview citation the way organic rank moves it for Product and Informational queries? We didn’t test GBP signals this round — we only noticed that organic rank mostly failed to explain Local citations. Next week’s arc tests the thing we think is actually driving them instead.

That closes the loop Monday’s fact-check opened and Tuesday’s pre-registration ran all the way through: the magnitude prediction held in a more extreme form than we bet, and both the intent and phrasing predictions turned out wrong — one by tying instead of trailing, one by reversing outright.

Frequently asked questions

Does Google’s classic AI Overview box cite by organic rank?

For Product, Informational, and How-to queries, yes, strongly: in our 24-query test, every citation we could match to a tracked organic position (85 of 85 in those three intents) sat in the top 10, none in positions 11-20. For Local-intent “near me” queries, the organic rank framework mostly didn’t apply — 11 of 13 citations in that bucket weren’t attributable to any ranked organic page at all.

Does this confirm the debunked “97% from top-20” stat?

No — it tests a different, narrower claim with real data instead of a fabricated citation. Monday’s fact-check found that specific 97%/top-20 figure doesn’t appear in the study it’s credited to. Our directly measured number, for top-10 specifically among citations we could rank-match, came in at 100%, which is in the same neighborhood but is a measured result from our own 24-query sample, not a confirmation of the original (nonexistent) source.

Why did the Local-intent prediction fail?

We predicted Local queries would trigger a classic AI Overview the least often of four intents. Instead, Local triggered in 6 of 6 queries, tied with Product and Informational and ahead of How-to (5 of 6) — the opposite of “least often,” which met our pre-registered reject condition directly. The deeper finding is that Local’s citations mostly weren’t drawn from the organic web index at all, appearing instead to source from Google’s own Business Profile or Maps-adjacent data.

Why did keyword phrasing trigger more Overviews than question phrasing?

We predicted the reverse, reasoning from a cited study’s finding that AI Overview-triggering queries tend to be longer and more complex. In our sample, all 12 keyword-phrased queries triggered an Overview versus 11 of 12 question-phrased ones — a single miss, so this should be read as evidence against our specific directional bet rather than a confirmed opposite effect.

What’s the next GEO Lab test?

Whether optimizing a Google Business Profile listing moves classic AI Overview citation for local-service queries, since this week’s data suggests organic web rank mostly isn’t what the Overview is drawing from for that intent.

Sources

• A Stat Says 97% of AI Overview Citations Come From Top-20 Pages. The Study It’s Credited To Doesn’t Say That. — Monday’s source check that opened this week’s arc.
• Does Google’s AI Overview Box Actually Cite by Rank? We Locked 3 Predictions and 24 Queries Before Testing It. — Tuesday’s pre-registered protocol and the three predictions scored above.
• AI Mode’s Source Count Fell 74% in Two Months. The Citations That Survived Were Mostly Ranked Top 3. — the AI Mode rank-retention data (46.2% vs. 8.3%) this week’s P1 estimate was built from.
• GeoParrot GEO Lab index — the running weekly arc this test belongs to.

GeoParrot is a GEO Lab: we test what AI search actually cites and recommends, then publish the method and the misses — including our own. This closes Week 20 of a running weekly arc. Next week: does optimizing a Google Business Profile listing move classic AI Overview citation for local-service queries the way organic rank moves it for everything else?