Optimizing for Google's AI Overviews

The results page has grown a mouth. For a widening set of queries, Google now leads with an AI Overview — a synthesized answer with a row of citations — and beyond Google, a growing share of questions get asked directly to Perplexity, ChatGPT, Claude, and whatever ships next quarter. Every one of these systems has the same two-part anatomy: a generated answer, and a set of sources it drew from and links to.

That second part is the game. You can't stop the answer from appearing, and you can't make it quote you by wishing. But which pages get retrieved, trusted, and cited is not random — it follows patterns you can influence with ordinary publishing decisions. Call it GEO, answer engine optimization, or just SEO with a new output format; the mechanics below are what actually moves it.

How these systems choose sources (the short version)

Whether it's an AI Overview or a Perplexity answer, the pipeline is roughly: take the query, run retrieval (which for Google leans heavily on the same index and ranking systems as classic search), pull the most relevant passages from the most trusted candidates, and generate an answer grounded in those passages, citing them. Three implications fall out immediately:

  • Classic SEO is the qualifying round. Pages that rank well for a topic are dramatically more likely to be in the retrieval set. There is no separate magic; if you're not competitive in ordinary search for a subject, you're mostly not in the AI answer either.
  • Selection happens at the passage level. The system isn't citing your page; it's citing a paragraph. A 3,000-word article with one crisp, self-contained, factual paragraph can beat a better article whose insights are smeared across ten meandering ones.
  • The systems favor sources that reduce their risk. Generated answers embarrass their makers when they're wrong. Sources with clear expertise signals, consistent factual framing, and specific attributable claims are safer to cite than vibes-based content — and the selection reflects that.

Write passages worth quoting

This is the highest-leverage change, and it's a writing discipline more than a technical one.

Lead with the answer, then elaborate

For every question your page addresses — every H2 that's phrased as or implies a question — the first sentence or two beneath it should be a direct, complete, standalone answer. Definition first, nuance after. "A staging site is a private copy of your website used to test changes before publishing them. Most managed WordPress hosts create one in a click; here's what that hides..." A retrieval system can lift that first sentence cleanly. It cannot lift an answer that begins three paragraphs of context later, and it often won't bother.

Make claims specific and attributable

"Backups are important" is unquotable filler. "A daily off-site backup with 30-day retention covers the two failure cases that actually happen to small sites: a bad update discovered late, and a compromised server" is a claim with edges — the kind of sentence that survives synthesis intact. Numbers, thresholds, named trade-offs, and rules of thumb (honestly framed as such) are citation magnets. So are stated opinions with reasoning; answer engines increasingly surface perspective, not just consensus.

Structure is retrieval infrastructure

  • Question-shaped H2/H3 headings that match how people actually ask.
  • Short, single-idea paragraphs — a passage retriever works with chunks, and a chunk containing one clear idea beats a chunk containing halves of two.
  • Lists and tables for anything enumerable or comparative. Tables get extracted into answers with striking frequency.
  • A summary or FAQ block for genuinely common questions — not as schema bait, but as clean answer-shaped text.

Publish what synthesis can't generate

Answer engines are consensus machines. They desperately need sources for things that aren't consensus: original data ("we measured load times across 40 sites"), first-hand experience ("after six months of running both..."), current facts (prices, versions, dates), and local knowledge. If your page just restates what ten other pages say, the engine can cite any of the ten. If it contains something that exists nowhere else, citing you is the only option.

The technical layer: be effortless to ingest

None of the above matters if the machines can't read you cleanly.

  • Server-rendered HTML is non-negotiable. Google's renderer handles JavaScript; most other AI crawlers do not. Content that only exists after client-side rendering is invisible to a meaningful share of answer engines. Standard WordPress themes pass this test by default; headless and heavily app-like frontends need SSR.
  • Decide your crawler policy deliberately. AI systems announce themselves in robots.txt-respecting user agents — GoogleBot's coverage feeds AI Overviews, and agents like GPTBot, ClaudeBot, and PerplexityBot feed the others. Blocking them is a legitimate business choice, but know what you're choosing: blocked crawlers can't cite you, and citations are referral traffic and brand exposure. For most small businesses, being findable everywhere beats a principled invisibility.
  • Structured data helps at the margins. Schema (Article, FAQPage, Product, LocalBusiness, and honest author markup) doesn't guarantee anything, but it disambiguates who you are, what the page is, and when it was updated — exactly the metadata a trust-weighting system wants. Cheap to add via any SEO plugin; add it.
  • Keep dates honest and visible. Answer engines weight freshness for volatile topics and display dates in citations. A real "last updated" (backed by real updates) beats both a missing date and a fake-bumped one.

Entity and reputation signals: being someone worth citing

Retrieval systems don't just rank pages; they model sources. Two sites can publish identical paragraphs and only one gets cited, because one of them is a known entity with a consistent footprint. Practical moves:

  • Consistent identity everywhere: same organization name, same author names with real bios, an about page that states plainly who you are and why you'd know. This is E-E-A-T plumbing, and answer engines drink from the same pipes.
  • Get mentioned where the models look. Citations breed citations: coverage in industry publications, expert quotes, podcast appearances, and active presence in the communities of your niche all raise the probability that retrieval systems associate your name with the topic. Notice this is just... reputation building. The tactics predate AI; the payoff surface is new.
  • Topical concentration beats topical sprawl. A site that publishes deeply on one subject builds a stronger source-topic association than a site that grazes twenty. If you want to be the citation for WordPress maintenance questions, a coherent body of work on exactly that is the asset.

Measuring something you can't always click

The reporting is genuinely awkward right now, but not hopeless:

  1. Search Console still works, squinting required. AI Overview impressions and clicks are folded into normal Search performance data. A pattern of rising impressions with falling CTR on informational queries usually means an AI answer moved in upstairs. Track CTR by query cohort quarterly.
  2. Watch referral traffic from answer engines. Perplexity, ChatGPT, Gemini, and Copilot all send identifiable referrals (check your analytics for their domains). Small numbers today, but the trend line is the point — and these visitors tend to arrive unusually well-qualified.
  3. Ask the engines yourself. Monthly, run your 10–20 money questions through AI Overviews, Perplexity, and ChatGPT with browsing. Note who gets cited. It's manual and unscientific, and it's also exactly how you find out which competitor's paragraph is beating yours, and why.
  4. Watch branded search volume. Citations build name recognition that surfaces later as people searching for you directly. It's the loosest metric here and possibly the most important.

What not to do

  • Don't mass-produce "answer-optimized" filler. A thousand thin Q&A pages generated to chase citations is the doorway-page playbook with new paint. The systems doing source selection are literally built to detect low-value synthesized text — publishing more of it at them is a strange plan.
  • Don't stuff FAQ schema onto everything. Marking up questions nobody asks doesn't create retrieval; it creates clutter.
  • Don't chase every fluctuation. AI Overview presence per query changes constantly as Google tunes thresholds. React to quarters, not days.
  • Don't abandon classic rankings. They're the front door to the retrieval set. Everything in this post assumes you keep doing the fundamentals.

The realistic posture

Here's the honest frame: answer engines are going to answer, with or without you. The choice available is whether your expertise is inside those answers — named, linked, and building your reputation — or outside them, perfectly preserved and unread. Earning citations turns out to require nothing exotic: rank for your subject, write answers a machine can lift and a human can trust, publish things only you could know, and be a consistent, identifiable source over time. Which is to say — the strategy for being quoted by machines is suspiciously identical to the strategy for earning people's trust. That convergence is the most reassuring fact in SEO right now.

Photo: Miguel Á. Padriñán on Pexels