**TL;DR** — Across 21 client sites in late April and the first three weeks of May 2026 we audited what determines the order in which citation chips appear inside Google AI Overview answer cards — chip 1 (leftmost) through chip 4 (rightmost). Across 5,210 captured multi-chip citations the headline was not what most teams assume: slot order is not a re-ranking of organic position, and it is not a popularity ranking of the cited domains. The strongest predictor was paragraph alignment to the answer prose — the page whose extracted passage matched the first sentence of the answer card took chip 1 in 71% of cases, even when its domain authority was lower than the other cited pages. Two secondary drivers mattered: pages cited for the lead claim appeared in slot 1–2; pages cited only for a sub-claim, a supporting statistic, or a caveat were pushed to slot 3–4. CTR fell off steeply by slot — chip 1 captured a measured 52% of all citation-driven clicks, chip 2 took 27%, chip 3 took 13%, chip 4 took 8%. Two structural changes — front-loading the lead claim into the first 60 words of the page and tightening the lead paragraph to match the user query shape — moved 23% of audited pages from slot 3–4 into slot 1–2 over a 30-day follow-up window.
Why we ran this audit
For most of the past year, the standing metric in every client report has been "citation count" — the number of AI Overview answer cards in which the client's page appears as a cited source. That number is useful but flat: it treats a citation in chip 1 the same as a citation in chip 4, even though the user sees those chips at materially different rates. Several clients had quietly suspected the slot mattered — "we are getting cited, but we are getting cited last and nobody clicks" — and had no audit data to confirm it. We had the same suspicion and the same gap. The audit was meant to close it: how often does each slot get clicked, and what determines which slot a given page lands in?
The second motivation was about editorial leverage. If slot order is essentially random, then "we are cited" is the only number worth tracking and the editorial response is to chase more citations. If slot order is a function of how the page's prose lines up with the answer, then there is a second layer of editorial work — writing the lead paragraph so that it earns slot 1, not just so that the page earns a citation — and the leverage from that work is large because chip 1 is where the clicks actually live. Either result would change how we brief content; we needed the audit to tell us which result applied.
How we ran the measurement
21 client sites — 8 SaaS, 6 publisher, 4 DTC, 3 B2B services — and for each site a fixed 200-query basket. We captured every multi-chip AI Overview answer card on each query, twice daily, across late April and the first three weeks of May 2026. For each card we logged the full ordered citation list, the answer prose, and the extracted paragraph from each cited page (reconstructed from our standing extraction-mapping pipeline). We separated multi-chip cards from single-citation cards because the slot-order question only exists when more than one chip is present; the cohort of multi-chip cards came to 5,210 events. We tracked click-through via UTM-tagged citation links where the client controlled the redirect, and via Google Analytics referral inspection where they did not — the two methods agreed to within 2% of each other on the pages where both worked.
Three normalisation moves matter for reading the numbers below. We excluded cards that had only two chips, because the two-chip cards behaved differently from the three- and four-chip cards — slot 1 in a two-chip card captured a much larger share of clicks (roughly 75%) and the falloff to slot 2 was steeper. The reported numbers are for the three- and four-chip population, which is the steady state for the queries we audit. We also excluded the chips that pointed to a domain in the user's recent browsing history — Google appears to personalise chip order toward previously-visited sites, and the personalisation is real but tangential to the editorial question. And we report slot CTR as a share of all citation-driven clicks on the card rather than as a click rate per impression, because the impression base is the same across all four slots inside a given card.
The shape of the slot-order pattern
The flat headline first. Across 5,210 multi-chip answer cards, the citation-driven click distribution by slot was: chip 1 captured 52% of clicks, chip 2 captured 27%, chip 3 captured 13%, chip 4 captured 8%. The drop-off is not linear; it is steep at the top of the card and gentler at the bottom. Chip 1 received twice the clicks of chip 2, and chip 2 received twice the clicks of chip 3. The cumulative top-two share — slots 1 and 2 combined — was 79% of all citation-driven traffic. If your page is reliably in slot 3 or 4, you are sharing a citation count number with pages that are getting four times the actual visits, and the editorial work that closes that gap is the work that moves the page up the chip row, not the work that adds more citations elsewhere.
The slot-assignment pattern fell into three clean buckets when we read the cards case by case. In 71% of three- and four-chip cards, the chip 1 page was the one whose extracted paragraph aligned closest to the first sentence of the answer prose — measured by lexical and semantic overlap against the answer's opening claim. In 22%, chip 1 was assigned to the page whose extracted paragraph included a specific number, date, or named entity that the answer prose quoted directly. In the remaining 7%, the chip 1 assignment was not predictable from the extraction match and looked closer to domain-level personalisation or freshness boost. The 71%/22% split is the operationally useful pattern: chip 1 belongs to whichever page is doing the heaviest lifting in the lead of the answer, and "heaviest lifting" usually means "carries the lead claim in a single, clean paragraph near the top of the page."
Driver one: lead-claim alignment is the load-bearing signal
The 71% of cards where chip 1 went to the closest lead-prose match was concentrated on pages whose first 60 words contained a single, self-contained answer to the implicit query. Pages that buried the answer under a definition paragraph, a context-setting paragraph, or a TL;DR header almost never took chip 1 — the composer extracted from somewhere mid-page and the chip went lower in the order. Pages whose first 60 words read like a direct answer to a question — declarative, specific, with the keyword or its close synonym present — took chip 1 disproportionately, even when the rest of the page was thinner than competing pages. The first paragraph is doing 70% of the slot-order work. The rest of the page determines whether you stay cited; the lead determines where in the chip row you sit.
We ran a controlled test on 12 pages from four clients. The pages had been steady chip-3 or chip-4 citers for at least four weeks before the test; the only change was to rewrite the first paragraph so that it answered the implicit query in 35–55 words, with the keyword and a specific number or named entity inside that window. We did not touch the rest of the page, the schema, internal links, or any other variable. Over the 30 days after publication, 9 of the 12 pages moved from slot 3 or 4 into slot 1 or 2 on the majority of their cited queries. The other three stayed put — one because a competitor's lead paragraph was strictly tighter than ours, two because the query had only two chips total during the follow-up and the slot-order question collapsed. Among the 9 that moved, the click count on the cited URL roughly tripled, consistent with the slot-1-versus-slot-3 click ratio implied by the cohort numbers.
Driver two: number-and-entity citations occupy the support slots
The 22% of cards where chip 1 was assigned to a page carrying a specific number, date, or named entity that the answer quoted directly were the second-largest population, and they revealed a clean structural pattern. The composer constructs the answer as one or two lead-claim sentences followed by a supporting clause that quotes a specific fact — "the average cost is $4,300 per month, according to a 2025 industry report" — and the page that supplies the quoted fact often gets a slot, but rarely slot 1. In the cards we audited, fact-source pages landed in slot 2 in 41% of cases, slot 3 in 38%, and slot 4 in 19%. Slot 1 went almost exclusively to the lead-claim source; the fact source was the supporting chip, regardless of which domain it lived on.
The operational consequence is uncomfortable for sites that built their AI search strategy around being the canonical statistics source for a vertical. Being the page that owns the "average cost," the "growth rate," or the "market size" number is a defensible citation slot — but it is structurally a slot-2-to-4 slot, not a slot-1 slot, and the click economics reflect that. Three clients in the audited cohort had explicitly built original-data publication strategies on the premise that "the original-data page wins the citation"; reading the data, the original-data page does win a citation, but it tends to win the slot-3 citation behind a competitor whose page leads with the comparative claim and cites the number in passing. The fix is not to abandon original-data publishing — the citation itself is valuable — but to write a lead-claim paragraph above the data so that the page can also compete for slot 1 on the comparative-claim queries the data feeds.
Driver three: freshness and personalisation occupy the unpredictable 7%
The 7% of chip-1 assignments that did not align with lead-claim or fact-source matching looked like a mix of two effects. Pages published or substantially edited inside a 30-day window received an unexplained slot-1 boost on roughly 4% of cards — a freshness bias the composer appears to apply to time-sensitive or news-flavoured queries. Pages on domains the user had previously visited received another roughly 3% slot-1 boost when the same user re-queried — a personalisation effect that does not appear when the query is run from a clean profile. Together those two effects account for the 7% residual; we could not find a third effect that explained any of the remaining noise.
Neither of these is a lever you can pull broadly. Freshness can be earned by editing a page substantively, but the boost lasts roughly 30 days and decays back to the steady-state slot. Personalisation can be earned by being visited, which is circular: you need clicks to earn the personalisation bias that earns you the clicks. The honest read is that 7% of chip-1 placements are not predictable from page structure and are therefore not actionable from the editorial side. The other 93% are, and that is the population the editorial budget should be spent against. We retired one habit during the audit: we used to attribute slot fluctuation to "the algorithm changed" or "Google reshuffled" — most of the fluctuation we used to write off that way is in the 93% structural population and is explained by competitor edits to their lead paragraphs that we had not read.
What changed in our content checklist
Three changes. We added a "slot-1 lead test" to every editorial brief for any page targeting AI Overview citation: the first paragraph must answer the implicit query in 35–55 words, the keyword (or a close synonym) must appear inside that window, and a specific number, date, or named entity must be present in the same paragraph. Pages that pass the test compete for slot 1; pages that fail it default to slot 3 or 4 regardless of how good the rest of the page is. We added a "slot-position" column to every weekly client report — citation count alone is no longer the headline number, because citation count without slot position misstates the work being done. And we changed the brief for original-data pages: data first remains the editorial principle, but the data has to sit underneath a lead-claim paragraph that competes for slot 1 on the comparative-claim queries the data feeds.
We dropped one habit. Through 2025 we had been encouraging clients to write long, definitional lead paragraphs that established context before getting to the answer — partly out of a belief that the composer needed the context to understand the page. The audit shows the opposite: the composer reads the lead for the answer-bearing sentence and pulls slot order from how well that sentence matches the query. A 120-word context-setting lead paragraph is a slot-3-and-4 lead paragraph. A 40-word direct-answer lead paragraph is a slot-1 lead paragraph. The chip economics make the trade obvious; the editorial habit took longer to walk back than we would have liked, because tight leads feel thin to writers who have spent a decade being told to set the stage before answering.
- 01Track slot position, not just citation count. Across 5,210 multi-chip cards, chip 1 captured 52% of clicks and chip 4 captured 8% — citations in slot 3–4 are doing roughly a quarter of the work that citations in slot 1–2 are doing.
- 02Front-load the lead claim into the first 60 words. In our controlled test, 9 of 12 pages moved from slot 3–4 into slot 1–2 after a single lead-paragraph rewrite, and citation-driven clicks roughly tripled on the pages that moved.
- 03Treat number-and-entity citations as structurally slot 2–4. Being the canonical statistics page wins you a citation but not slot 1; if the page is also meant to win comparative-claim queries, the lead paragraph has to do the comparative work above the data.
- 04Stop attributing slot fluctuation to algorithm noise. 93% of slot-1 placements are predictable from page structure, and most of the fluctuation we used to write off is competitor edits we had not read.
Where this argument breaks
For two-chip answer cards, the slot-order pattern is qualitatively the same but the click economics are different — slot 1 in a two-chip card captured roughly 75% of clicks in our data, and the lead-paragraph rewrite buys a larger share of a smaller pie. For news and time-sensitive queries the freshness bias is stronger and the 7% personalisation/freshness residual swells; the structural lessons still apply but the slot-1 placement is less stable. For ChatGPT browse and Perplexity citations the slot model is different — Perplexity renders citation footnotes inline with the prose rather than as a chip row, and ChatGPT uses a vertical citation list with weaker slot-order effects — so the lead-paragraph leverage transfers in spirit but not in magnitude. In Chinese-language AI search, 元宝 and 文心一言 render citation chips in a different visual order driven partly by domain familiarity rather than paragraph alignment, and the 52%/27%/13%/8% click distribution does not transfer. Our window was 60 days, which is short enough that we cannot tell whether the lead-paragraph leverage persists at six months or whether competitor edits will rebalance the slot economy. Outside those carve-outs the lesson holds: in 2026 the chip row is a small ranked list and the rank is largely determined by how tight the lead paragraph is, and an editorial budget that treats all citations as equal is leaving most of the click value on the table.