Eight of the ten commercial queries now fire an Overview. Our URL holds position two in the classic list and does not appear in the citation stack.
That gap is the job. Eligibility is whether Google’s systems can treat the page as a source. Control is whether they pick it for that query, on that day, in that language, for that device. We work the first. We do not own the second.
If you are the brand lead chasing AI Overview citations, stop treating the block as a ranking we can buy with more words. Treat it as a source-selection problem with a hard ceiling.
Eligibility is a filter, not a lever
Google AI Overviews do not run on every query. When they do run, they pull from pages the systems already trust enough to extract. That is eligibility: crawlable, indexable, clear, and close enough to the question that a model can lift a claim without inventing one.
Eligibility is binary-ish. Either the page can be used or it cannot. Thin source clarity, blocked snippets, contradictory H2s, or a canonical that points at a weaker URL will knock you out before anyone talks about “AI Overview SEO”.
We check eligibility before we talk about citations. Index coverage in Search Console. Rendered HTML, not the CMS preview. The exact question the Overview is answering, not the keyword we wished we ranked for.
A page can be eligible and still never be cited. That is not a reporting bug. That is how the feature works.
Control ends at source clarity
Control, in this work, stops at making the page easy to extract and hard to misread. We do not choose the three links under the generated paragraph. We do not choose whether the Overview appears at all.
Source clarity is the part we can actually ship. One claim per section. Dates on anything that ages. Named entities spelled the same way the SERP already uses them. Answers that stand without the H1. Tables that survive a copy-paste into a notes app.
When two of our URLs compete for the same extraction, we lose control we never had. The systems pick one, or pick neither, and cite a forum thread instead. Consolidate the claim. Kill the duplicate. Leave one eligible source.
Anyone selling a switch that “gets you into google ai overviews” is selling a ranking they cannot operate. We will not.
What Search Console actually records
Search Console will show impressions, clicks, and average position for queries that also fire Overviews. It will not label the citation. It will not tell you the page was used as a source. It will not split Overview traffic from classic blue-link traffic in a way we can trust for a board slide.
We still use it. Impression spikes on question-shaped queries are a hint the Overview is running. Click-through collapse on the same queries is another. Neither proves we were cited. Neither proves we were eligible.
The only honest check is still the SERP. Incognito is not enough. Logged-out, target locale, the device the brand actually cares about. Screenshot the citation stack. Note which domain, which URL, which claim got lifted. Do it on a schedule, not when someone in Slack is angry.
If Search Console shows the URL as not indexed, eligibility is already dead. Fix that before you rewrite the intro.
How google ai overviews choose sources
Google’s public description of the feature lives on the Google AI Overviews help page, and it is deliberately thin. Overviews appear when systems decide they can be helpful. They include links to supporting web content. They are not a separate index we submit to.
That page is worth reading once, then ignoring as a playbook. It will not tell you the scoring. It will not list the eligibility tests. It will not explain why a Reddit comment beat a 2,000-word guide last Tuesday.
What we can observe in live SERPs is consistent enough to work from. Overviews prefer pages that answer the question in the first extractable block. They mix source types: publishers, official docs, forums, our own category of commercial pages. They change. A citation this morning is not a citation next week.
AI overview SEO, as a phrase, is just ordinary SEO with a tighter extraction test. If the paragraph cannot stand alone, it will not be lifted. If the claim is hedged into mush, a blunter competitor gets the slot.
Being used is not the same as being cited
A page can inform the generated text and never earn a link in the stack. We have no reliable way to prove that from the outside. Do not build a KPI on “we were used”. Build it on visible citations, and treat even those as unstable.
Visible citation is the only outcome a brand lead can take to a quarterly review without looking foolish. Screenshot, URL, query, date, locale. That is the artefact. Everything else is a story.
If the Overview quotes our unique phrasing and cites someone else, we still do not get the click. Source clarity helps extraction. It does not assign credit. That is the control problem again.
We log citations. We do not forecast them.
Queries worth mapping first
Not every query deserves this work. Informational questions with a stable answer are where Overviews cluster. Commercial-investigation queries sometimes fire them, sometimes do not. Navigational queries almost never should.
Map the cluster that already sends impressions in Search Console. Separate the questions the Overview is actually answering from the head terms on the media plan. The Overview for “how long does X take in the Philippines” is a different extraction problem from “best X agency”.
We run that split through keyword research services before anyone touches a template. The output we need is not a volume dump. It is a list of Overview-triggering questions, the current citation stack, and whether our eligible URL even targets the question in a form a model can lift.
If the money query never fires an Overview in the locales we care about, stop spending the sprint on citations. Spend it on the classic result we can still influence.
On-page work that extraction can follow
The useful on-page work is boring. Direct answer in the first two sentences of the section that matches the question. No throat-clearing. No “in today’s digital landscape”. The H2 should be the question or a close paraphrase. The paragraph under it should be the answer.
Lists help when the question is a process. Tables help when the question is a comparison. Neither helps if the cells are marketing. Put the number, the constraint, the exception.
Schema is not a citation cheat. FAQ markup that repeats fluff will not make an ineligible page eligible. We still mark up what is true and visible. We do not mark up a theatre version of the page.
Internal links should point at the one URL we want extracted for that claim. If five blog posts all half-answer the same question, we have five weak sources and no control.
Update dates only when the substance changed. A fake freshness stamp is source clarity going backwards.
Technical blocks that kill eligibility
Before content, we check the boring failures. `noindex`. `nosnippet`. A robots rule that looked temporary in 2022. JavaScript that never renders the answer for Googlebot. A canonical to a thinner parent. Soft 404s on the URL Search Console still thinks is live.
Migrations are where eligibility dies quietly. The Overview keeps citing the old URL until it 404s, then it cites someone else. Redirects do not transfer a citation. They transfer a chance to become eligible again.
Hreflang mistakes in PH/EN setups send the wrong language into the wrong Overview. If the citation stack is in a language we did not write, we are not in the game. Fix the locale before we rewrite the English.
We treat these as eligibility bugs, not “AI issues”. The Overview cannot cite a page the index will not use.
Reporting without a fake KPI
A brand lead needs a report that survives a sceptical CFO. We do not put “Overview share of voice” on a slide unless we defined the query set, the locale, the device, and the capture method.
Useful reporting looks like this. Query set of twenty. Weekly capture. Cited or not. Which URL, ours or not. Impression and CTR from Search Console on the same queries, with the caveat that the two systems do not join cleanly. Notes when Google changed the layout.
Park that watch where the rest of the SEO work already lives. The seo command centre is where we keep eligibility, citations, and classic rankings in one view so the citation chase does not become a side spreadsheet nobody trusts.
If the only movement this quarter is more Overviews firing and fewer clicks, that is a market change, not a content failure. Say so. Do not dress it up as a win because we “almost” got cited.
When chasing a citation wastes the quarter
Some clusters will never cite a commercial site. Official regulators, encyclopedias, and large forums occupy the stack and will keep occupying it. Eligibility work on our URL is still useful for classic results. Citation work is not.
If the Overview answer is a definition we cannot uniquely own, we are competing with every publisher who wrote the same sentence. Spend the sprint on a query where we have a primary source: original data, a method, a constraint the forums do not have.
If the brand needs a citation for a board narrative and the SERP will not give one, say that early. The next useful thing is not another rewrite. It is a different query, a different URL, or a stop.
We will not promise a citation. We will tell you whether the page is eligible, whether the query even fires an Overview, and what source clarity still missing.
Decide the next useful step
Open Search Console. Export the question-shaped queries with impressions. Check ten of them live, in the locale that pays the bills. Note which fire an Overview and which of those cite us.
If the URL is not indexed, that is the only job this week. If it is indexed and the answer is buried under a brand story, that is the on-page job. If the answer is clear and we are still not in the stack, we are in the control gap. Do not rewrite for the sake of motion.
If the query set is a mess, we map it properly, then we audit eligibility on the URLs that could actually be sources. If that fit is real, ask for a scoped audit on the cluster, not a site-wide essay about AI.
Citations are a lagging, unstable signal. Eligibility and source clarity are the work. Control is the part we do not get. Plan the quarter around the first two and you will stop being surprised by the third.
