"What do we need to do to rank in AI Overviews?"
I get that question on nearly every discovery call now, usually from someone who has already been quoted a five-figure retainer to answer it. And the honest response makes people visibly uncomfortable: nobody outside Google knows, because Google has never published a ranking algorithm for AI Overviews. Not a document, not a factor list, nothing.
That does not mean you're helpless. It means the work that actually moves the needle is less glamorous than what's being sold, and considerably cheaper. Let's go through what the evidence supports, in order.
First, the damage is real
Start with the number that made everyone panic. Pew Research analyzed 68,879 real Google searches in March 2025. When the results page carried an AI summary, users clicked through to a result on 8% of visits. Without a summary, 15%. Clicks on links inside the summary itself: 1%.
Ahrefs put a figure on the top spot specifically. They found AI Overviews correlate with a 34.5% drop in click-through rate for position one between 2024 and 2025. Stretch the window from December 2023 to December 2025 and the decline reaches 58%.
Correlation, to be clear: results pages changed in a dozen ways over those two years, so nobody has isolated AI Overviews as the single cause. But if your CTR chart bends downward while your rankings hold steady, you're not imagining it.
Google's counterpoint, and what it's worth
Google's own documentation on AI features makes a quality argument rather than a volume one: clicks from AI Overviews are "higher quality (users more likely to spend more time on the site)."
I think that's probably true, and I also think it's conveniently unfalsifiable from the outside, because Google publishes no click volume next to it. My read is that both things are happening at once: fewer clicks arrive, and the ones that do have already read a summary and decided to keep going anyway.
Whether that trade works depends on your business model. If you sell display impressions, losing half your clicks for better-qualified survivors is a disaster. If you sell a $60,000 integration project, ten pre-qualified visitors beat a hundred tire-kickers. Check your own conversion data before deciding which camp you're in.
Nobody has published a ranking algorithm for AI Overviews
This is the part the industry keeps skipping, so let me be blunt about it. Any list of "AI Overview ranking factors" you've been shown was invented by whoever showed it to you. There's no confirmed mechanism and no published weighting.
What Google has published is a direction. From its AI optimization guide: "optimizing for generative AI search is optimizing for the search experience, and thus still SEO." The mythbusting section of the same guide is even more direct about the new file formats people are being sold: "You don't need to create new machine readable files... as Google Search itself doesn't use them," and doing so "will neither harm nor help your site's visibility or rankings in Google Search." The guide also states that structured data "isn't required for generative AI search."
Read that carefully and a lot of invoices stop making sense. There is no second algorithm to game and no new artifact to buy your way into. If the terminology around all this has you confused, we untangled it in AEO vs GEO vs SEO, but the short version is that most of the new acronyms describe the same work.
So what does influence whether you get pulled into a summary? Four things, ranked by how much evidence sits behind them.
1. Boring SEO fundamentals still decide most of it
An AI Overview is assembled from pages Google can find, crawl, and already considers relevant to the query. If you're not in the index, you're not a candidate. If you rank on page four, you're a weak candidate.
So the unsexy checklist governs: crawlable pages, sensible internal links, a sitemap that reflects reality, topical depth on the queries you care about, and enough authority that Google trusts the page. Nothing here is new. That's the point, and it's why moving budget from "AI Overview optimization" back into ordinary SEO is usually right.
2. Write in answer shapes
Here's professional judgment rather than a Google-published factor, and I'll flag it as such: content structured as discrete question-and-answer units is easier to lift into a summary than the same information buried in a 400-word narrative.
What that looks like in practice:
- Use the real question as the heading, phrased the way a person types it, not as a keyword stub.
- Put a direct answer in the first 40 to 60 words underneath, before the context and the caveats.
- Make each passage survive extraction. If you cut one paragraph out of the page and hand it to a stranger, does it still make sense on its own, or does it collapse without the three paragraphs above it?
- Use tables for comparisons, prices, and specification lists. Tables are unambiguous in a way prose never is.
- Attach dates and numbers to claims. "Pew, March 2025, 68,879 searches" is quotable. "Studies show" is not.
Costs nothing, and impatient humans skim exactly the same way. It's the argument we made years ago about user experience in mobile app development: clarity for the hurried user turns out to be clarity for everyone.
3. Entity clarity and schema, honestly framed
Google says structured data isn't required for generative AI search. I believe it, and I still ship JSON-LD, for one specific reason: disambiguation.
Say you run a fictional analytics company called Northwind. So does a wine importer and a school district. Organization markup with a `sameAs` set, consistent naming across your site, an author entity matching your bylines, and a Service entity that spells out what you sell all help a machine work out which Northwind it's reading. Vercel's Agent Readability Spec (March 2026) puts JSON-LD with `dateModified` plus a `BreadcrumbList` on its per-page checklist, which is a fine baseline.
Treat it as a one-time task, though, not a program. If someone is billing you monthly to maintain schema on a fifty-page site, that's a line item to question.
4. Rendering and access hygiene, where the real failures live
This is the category nobody buys and almost everybody needs.
Vercel's crawler analysis (November 2024 data) found that the major AI crawlers operated by OpenAI, Anthropic, Meta, and Perplexity fetch JavaScript files but never execute them. If your content only exists after hydration, those systems see an empty shell. The same analysis found AI crawlers waste roughly a third of their requests on 404s (34.8% for ChatGPT's crawler, 34.2% for Claude's, against 8.2% for Googlebot), which is what happens when a crawler has to guess at URLs.
For Google specifically the rule is simpler: if Googlebot can't render and index the page, it can't appear anywhere in Google, AI Overviews included. Server-side rendering is not an AI trick, it's the oldest advice in the book, and it's still the single most common thing I find broken.
Then there's markup that says what things are. Google's web.dev guidance on agent-friendly sites describes the accessibility tree as "a high-fidelity map that ignores the visual 'noise' of CSS." A `<div>` with a click handler doesn't appear on that map. Neither does an unlabeled input. Google is refreshingly relaxed here (its own words are "don't worry about perfect code"), and its closing line is the one I quote to skeptical CTOs: "Everything we suggest to make a site 'agent-ready' also makes sites better for humans."
Worth separating one thing, because the two get conflated constantly. Assistant-specific bot rules are about assistants, not about Google. OpenAI documents that sites opted out of OAI-SearchBot "will not be shown in ChatGPT search answers," and John Mueller's warning (reported by Search Engine Journal, June 2026) applies broadly: "Don't block agents. I think that hurdle will be the biggest, for most sites." Also remember robots.txt is a preference rather than a fence. TollBit measured around 15% of AI scrapes ignoring an active disallow in the first half of 2026, with ChatGPT-User at 54%. Enforcement lives at your CDN or WAF, if you want it at all.
How to actually see whether you're in AI Overviews
You can't manage this without a baseline, and the baseline is manual. It's also free.
Build a list of 20 to 30 real queries, written the way buyers type them mid-problem rather than the way keyword tools phrase them. Run each one in a clean, logged-out browser and record four columns: did an AI Overview appear at all, were you mentioned, which domains got cited, and which competitor took the slot you wanted.
Then repeat the same list monthly. AI Overviews don't trigger on every query and the output varies between runs, so treat each pass as a sample rather than a verdict. The cited-domains column is the most useful thing you'll build: it tells you which third-party pages Google trusts in your category, which is your outreach and PR plan for the next quarter.
Alongside that, two machine-readability checks take about ten minutes:
https://r.jina.ai/https://yoursite.com/your-best-page npx lighthouse https://yoursite.com --only-categories=agentic-browsing
Jina Reader returns your page as plain text, roughly what a non-rendering fetcher gets. If it comes back as navigation with no body copy, stop reading articles like this one and go fix your rendering. Lighthouse's agentic browsing category is scoped to browser agents rather than AI Overviews, but it catches accessibility-tree and layout-stability problems cheaply. We ran it on ourselves and scored 2/3 at the time, failing on our own llms.txt formatting, which we fixed to reach 3/3. Everyone fails something on the first run.
What to stop paying for
Guaranteed AI Overview placement, first. Nobody can guarantee it. We can't, and neither can anyone who quotes you a price for it, because there's no published mechanism to guarantee against.
Content rewrites of pages a crawler can't read, second. That's the most expensive mistake in the category, and we broke down why in LLM SEO: what actually gets you cited.
Third, llms.txt sold as an AI Overviews lever. Google's documentation says Search ignores these files outright, and Ahrefs looked at 137,210 domains to find that 97% of llms.txt files received zero traffic in May 2026. Ship one if your developer audience wants it. Don't buy one to rank.
And anything justified by a leaked ranking factor. There isn't one.
Where to start on Monday
Open a spreadsheet, write down 20 queries, and run them. That's your baseline, and until it exists every decision you make here is a guess. Then run the Jina Reader check on your three best pages, because if a fetcher sees an empty shell, nothing else on this list matters yet.
After that, the sequence is dull and effective: fix rendering, fix markup semantics, restructure your top pages into answer shapes, ship schema once, and keep doing ordinary SEO. Google's own position is that this is still SEO. I'd take them at their word and spend accordingly.
If you'd rather have someone else find the failures first, we run a free AI SEO audit: we fetch your pages the way the crawlers do, check your bot rules and rendering, review your top pages for extractable answers, and send back a ranked list of fixes. Findings, not a pitch deck.
