How to rank in AI Overviews: the 2026 playbook

Why 'ranking' in AI Overviews is the wrong frame

An AI Overview is a synthesised answer that draws from a curated set of source pages — usually three to five. Google (and increasingly Bing, Perplexity, and ChatGPT search) shows those sources in a small sources drawer beside or below the answer. That drawer is the new page-one real estate. A #1 organic position sitting below an Overview that already answered the question routinely loses 30 to 55 percent of its clicks; a citation slot inside the Overview holds its click-through in the 8 to 14 percent range for transactional queries, according to what we track across roughly 200 client queries per week.

So the question is not 'how do I rank higher'. It is: 'how do I become one of the three-to-five pages the model trusts to cite for this query?' The moves you run to answer that question overlap with classical SEO in about 40 percent of the surface area — technical health, crawlability, internal linking — and diverge sharply in the other 60 percent. That divergence is where most 2024-era SEO playbooks quietly stop working.

The other reason to reframe: 'ranking' implies a single number you can chase. Citation eligibility is a compound of four axes we track (answer clarity, entity strength, topical placement, freshness). Optimising one axis in isolation almost never earns a citation. Optimising all four together is what compounds.

The three signals that actually earn citations

We reverse-engineered citations across roughly 4,200 tracked queries in the first half of 2026. Three signals dominated, in this order.

Signal one: the source page directly answered the exact question in the first 60 to 120 words of body copy. Not 'the answer is eventually on the page'. Not 'the answer is implied by the H2 structure'. The literal sentence the model can lift, above the fold, in plain prose. Pages that bury the answer under a definition, a history section, or an intro that markets the company almost never get cited — even when they rank organically for the same query.

Signal two: unambiguous entity signals. Named authors with Person schema and sameAs to real profiles. Organization schema on every page with a consistent NAP (name, address, phone) that matches the Google Business Profile character-for-character. Internal links using clean, descriptive anchor text rather than 'click here' or 'learn more'. The models trust pages that make their identity — and the identity of the people writing them — machine-legible.

Signal three: topical placement inside a cluster. A brilliant one-off article on a topic scores far worse than a good pillar page supported by twelve directly linked sub-articles. Cluster geometry — a hub with descriptive spokes — is what tells the model 'this site actually knows this domain'. We've watched 40-page sites out-cite 40,000-page publishers on diagnostic queries because the small site had the topic-cluster geometry right and the large site didn't.

A fourth signal is emerging and we now weight it seriously: visible freshness with a real content diff. A 'Last updated' line in body copy paired with genuinely refreshed statistics, pricing, or recommendations extends citation persistence through model updates by roughly 2.3× in our data. Ghost-updating the date field without changing anything does not — the models appear to check.

The first-100-words rewrite

This is the single highest-leverage move on the entire playbook and it's the one most teams skip. Pull your top ten organic-traffic pages. For each, rewrite the first 100 words so that the literal question the page targets is answered — in one clean sentence — inside the first two sentences of body copy. Follow it with 60 to 80 words of specifics: the pricing, the timeline, the edge case, the local nuance, the number.

The reason this works: users clicking through from an AI Overview have already been oriented by the summary. They don't need a primer; they need the specifics the summary couldn't cover. Serving them a primer they've already read bounces them straight back to the SERP. Meanwhile the models pick up the clean answer sentence as citation-eligible copy.

A rewrite checklist we tape above the desk: does sentence one answer the exact query? Does sentence two add a specific — a number, a timeline, a location — that a summary can't? Does paragraph two show something a summary can't, like a comparison table, a screenshot, or a live quote? If all three answers are yes, the page tends to hold click-through even as AI Overviews expand around it.

In production, this move alone lifted average citation rate on rewritten pages from 4 percent to 19 percent across a sample of 62 pages we've re-tested at 90 days. Nothing else on this list moves the number that fast on that little effort.

The entity graph that machines trust

Every citation-earning site we've analysed has an entity graph that a machine can read without ambiguity. The graph is small and it's stable: an Organization with sameAs pointing to LinkedIn, Wikidata, and Google Business Profile; named Persons for authors with their own bylines, headshots, and Person schema; and Product or Service entities where relevant, each with clear identifiers.

The mistake most sites make is treating the About page and Author pages as marketing copy. In AI-search terms they are the primary trust anchor. An About page with no Organization schema, no founder photo, no press or credential proof is invisible to the models the same way an anonymous forum post is invisible. Fix this and every article on the site starts benefiting from the entity halo.

For local businesses, the highest-leverage entity move is aligning on-site NAP with Google Business Profile character-for-character. Every mismatched suite number, abbreviated street type, or format-inconsistent phone number weakens the entity signal. This costs nothing and takes an afternoon; skipping it invalidates half the schema work you'll do afterwards.

The schema stack — what pulls its weight

Not all schema pays off. We run four types on client sites and each earns its place: Organization on every page for entity trust, Person for author trust on articles, Article/BlogPosting on editorial content for freshness and byline attribution, and FAQPage where the page genuinely answers a small set of distinct questions. For local clients, LocalBusiness (with the correct Google Business Profile handle in sameAs) and Service schema on service pages both move the needle for local citation eligibility.

Types we don't bother with unless there's a specific reason: HowTo (deprecated by Google for most surfaces), Speakable (limited support), and the various niche subtypes of CreativeWork. The rule we use: if adding schema doesn't make the page-model relationship less ambiguous, it's decoration and it invites markup-error warnings without upside.

Ship schema in JSON-LD in the head; never inline microdata that fights with the visible HTML. Validate every template with the Rich Results Test after each deploy — one broken JSON-LD block on the site template silently invalidates the schema across every page it renders on. We've seen a client lose citation share for four weeks from a single trailing-comma error in a template.

Tracking citation share weekly

You cannot manage citation share if you cannot see it. We run a lightweight tracker on 200 to 500 target queries per client, weekly, using a headless-browser script that captures whether an AI Overview appeared, which sources it cited, and whether the client site was one of them. Results land in Looker Studio alongside classical ranking data. The reporting cadence matches the ranking dashboard; the KPI is 'citation share on tracked queries', which we treat as a first-class metric.

Commercial GEO trackers exist now (Peec, Otterly, Athena, Profound). Most are early, most cost more than the DIY stack, and none we've tested provides materially better signal for our use case. If you value the visualisation and are past 500 queries, the paid tools become reasonable. Under 500 queries, the DIY stack is fine and much cheaper.

The reason to track weekly rather than monthly: model updates roll out unevenly and citations churn faster than organic positions. Losing a citation on a high-value query and finding out three months later is the failure mode. Losing it on Tuesday and re-earning it by the following Tuesday is normal operations.

The 30-day sprint we run for new clients

Week one: audit the entity graph, fix Organization and Person schema across all templates, align NAP with Google Business Profile, and identify the top 20 pages by organic traffic. Ship an initial citation tracker over the top 200 queries for the domain.

Week two: rewrite the first 100 words of the top ten pages. Add FAQPage schema where the page genuinely answers a set of distinct questions. Add clean, descriptive internal links from adjacent pages into each rewritten page.

Week three: publish two supporting articles per top-priority topic to strengthen the cluster geometry around the pages we care about. Refresh the top ten pages' publication dates only if we actually changed something substantive in body copy — never as a cosmetic update.

Week four: measure. Pull the first week of citation-share data, compare it to baseline, and decide which cluster to strengthen next. In our engagements this sprint typically shifts citation share on tracked queries from 8 to 12 percent baseline to 18 to 24 percent by day 30, with the full compound landing at 60 to 90 days as freshness signals mature.

If you want us to run this sprint on your site, that's what an initial engagement looks like — the same playbook, on your queries, with weekly reporting from day one.