Short answer. Vet a Google AI Overviews partner on four things: measurement rigour (AI Overviews are volatile by query, location, device and session — a partner sampling once per query is guessing), content provenance (can they show who wrote and checked a claim, and against what source), multilingual execution (can they publish answer-ready content in your markets, not just report on them), and answer-ready delivery (do they produce passages an engine can lift intact, or blog posts with a keyword in the title?). Anyone guaranteeing placement in AI Overviews is selling something they do not control.
Google AI Overviews changed the economics of a search result. A query that used to send a click now often resolves in the answer panel, and the brands named inside it inherit the attention that ten blue links used to distribute. The work of getting there is real, learnable and — critically — different from both classical SEO and from optimising for ChatGPT, because AI Overviews are grounded in Google's own index.
That last point is the one most proposals get wrong, in both directions. Partners who treat AI Overviews as "just SEO" ignore the passage-level selection that decides which page gets quoted. Partners who treat it as a wholly new discipline sell you an "AI-first" content programme on a site that Google cannot crawl properly. This guide is how to tell those apart before you sign.
How do Google AI Overviews actually select sources?
Enough mechanics to evaluate a proposal:
- AI Overviews are grounded in Google Search. There is no separate AI Overviews index to be admitted to. If a page is not indexed and reasonably retrievable for the query, it cannot be drawn on. Classical SEO fundamentals therefore remain a precondition, not a legacy concern.
- Selection happens at passage level, not page level. The unit that gets used is a self-contained span of text that answers the question. A page that ranks well but buries the answer in paragraph nine of a narrative competes badly against a page that answers in its first two sentences under a matching heading.
- A query expands into several sub-questions. A single user query is decomposed, and different sources may supply different parts of the answer. This is why a page can appear for a question it does not obviously target, and why coverage is better thought of as covering a topic's sub-questions than a single keyword.
- Presence is unstable. Whether an AI Overview appears at all varies by query, query intent, location, device, language and over time. Any measurement built on a single check per query will report movement that is really variance.
- Reporting is limited by design. Google surfaces AI Overview impressions and clicks inside standard Search Console performance data rather than as a separately labelled breakdown. A partner claiming to read your exact AI Overview click share out of Search Console should be asked precisely which report they mean — and any answer that implies a dedicated AIO dimension deserves a second look against Google's current documentation.
Two direct implications for vetting: a partner who cannot explain passage-level selection will produce pages that rank and do not get quoted; and a partner whose measurement is not repeated sampling will report noise as progress.
What does rigorous AI Overviews measurement look like?
Five requirements. A partner meeting all five is rare and worth paying for.
- A fixed query set, defined before work starts. Category questions, comparison questions, and branded questions, held constant across periods.
- Repeated sampling. Several checks per query per period, across at least the locations and device types that matter commercially. Report a presence rate, not a yes/no:
AI Overview presence rate = Checks where an AI Overview appeared ÷ Total checks for that query
Citation share = Checks where your domain was cited ÷ Checks where an AI Overview appeared
- Location and language segmentation. A national average is close to useless for a brand selling in several markets. Segment by market, and by language where they differ.
- A pre-work baseline. Without it, nothing later can be attributed. This is the single most common gap in AI-visibility proposals and the cheapest to close.
- Raw evidence retained. Screenshots or captured text of the overview, timestamped, with the query and locale. Ask to see a period's raw file for an existing client, redacted. A partner who can only produce a dashboard cannot show you what it was computed from.
Add one honest limitation to any partner's claims: correlation between their work and a citation is rarely clean, because Google's own systems change underneath the measurement. A partner who acknowledges this is more trustworthy than one who presents a clean attribution chart.
Why does content provenance matter for AI Overviews?
Because the exposure profile changes when your text is quoted rather than linked. A claim inside a page is read by a visitor in context. The same claim lifted into an AI Overview is read as a stated fact, attributed to your brand, by someone who never saw the page.
Three practical requirements:
- A fact-check standard. Which claims are verified against a source and which pass on the writer's judgement. If a partner has no standard, they have no answer when a quoted claim turns out to be wrong.
- A record per asset. Who wrote it, who reviewed it, when, and which sources were consulted. If AI assistance was used in drafting, the record should say so and at what editorial level a human intervened.
- Currency management. Quoted statistics age. A partner should have a review cadence for dated claims, and a way to find every asset containing a superseded figure.
Red flag: a partner proposing high-volume, thinly-sourced content to "increase surface area". Evidence density, not volume, is what the published research supports — in the ACM KDD 2024 benchmark across 10,000 queries, adding statistics raised citation visibility by up to 40% and authoritative quotations by roughly 30%, while keyword stuffing scored −10%.
What does multilingual execution require?
If you sell in more than one language, AI Overviews performance is a per-market question. Two markets speaking the same language will return different overviews, different competitor sets and different question phrasing.
Require, per market:
- Query sets authored by native speakers, not translated from English.
- In-market content production, not translated English pages. The translated page answers the English question in another language.
- Hreflang and per-language canonical correctness. This is boring and it gates everything: mis-declared language alternates cause the wrong market's page to be indexed and quoted.
- Native-speaker review of anything making a claim, because claim legality varies by market and a translation pass will not catch it.
The test question is blunt: "How many in-market native speakers do you have who can write, not just review, in each of our languages?" Supported-language counts answer a different question.
What is "answer-ready" content, concretely?
A page is answer-ready when a machine can lift a correct, attributable passage out of it without the surrounding context. In practice:
| Property | What it means |
|---|---|
| Question as a literal heading | The heading matches how buyers phrase the question, not an internal product name |
| Answer in the first two sentences | The direct answer leads; the elaboration follows |
| Self-contained passages | Each passage makes sense quoted alone — no unresolved "as mentioned above" |
| Evidence density | Statistics, formulas, named sources, dates — with a source behind each |
| Definitions stated plainly | One-sentence definitions an engine can quote verbatim |
| Structured data | Correct, non-contradictory schema; FAQ and Article markup that matches the visible page |
| Freshness signals | Real dates that reflect real updates, not a rolling timestamp |
| Crawlable without JavaScript | The answer exists in the served HTML, not only after hydration |
That last row disqualifies a surprising number of otherwise good sites. Ask any partner to show you what a crawler receives with JavaScript disabled, on your most important page, before they propose content work.
How should you score partners?
| Criterion | Weight | Evidence to require |
|---|---|---|
| Measurement rigour | 30% | Baseline procedure, sampling frequency, raw run file, locale segmentation |
| Answer-ready delivery | 25% | Two published examples they produced, and the passage that got cited |
| Content provenance | 15% | Fact-check standard, per-asset record, currency review cadence |
| Multilingual execution | 15% | In-market writer counts per language |
| Technical SEO foundation | 10% | Crawl and rendering audit before content scope is set |
| Reporting and ownership | 5% | Data export, content ownership at contract end |
Disqualifying conditions: any guarantee of AI Overview placement; no pre-work baseline; single-sample measurement; a proposal that begins with content volume before a crawl and rendering check; claims that a special file or schema type grants AI Overview inclusion.
On that last one — be specific in the meeting. There is no "AI Overviews schema". And llms.txt, whatever its merits for other AI systems, is not used by Google Search; a partner selling it as the route into AI Overviews is either uninformed or hoping you are.
What should you ask a Google AI Overviews partner?
- Walk me through how a passage gets selected for an AI Overview.
- Show me a raw measurement file from a client period, redacted.
- How many samples per query per period, and across which locales and devices?
- What is your baseline procedure, and what do you do if we insist on skipping it?
- Show me a page you produced and the exact passage that was cited.
- What is your fact-check standard, and who applies it?
- What does a crawler receive from our top page today with JavaScript disabled?
- Which of our queries do you expect not to move, and why?
- How many in-market writers do you have for [hardest market in scope]?
- What has Google changed in the last year that invalidated something you previously recommended?
Question 8 and question 10 do most of the work. A partner who expects everything to move is not being straight with you, and one who has never had a recommendation invalidated has not been paying attention.
How Lifewood approaches this
Lifewood treats AI Overviews as one surface inside a single AEO and GEO programme rather than a standalone product, because the underlying work — entity consistency, answer-ready passages, evidence density, crawlable delivery — is shared across engines, while the measurement is engine-specific.
The measurement instrument is operated in-house with fixed prompt and query sets and a pre-work baseline, and results are reported per market rather than as a single global figure. On execution, 50+ languages, 40+ delivery centres across 30+ countries and 56,788 contributors mean in-market authorship rather than translated English, including in markets most providers cover with machine translation. Lifewood's AI-data heritage runs to 2004 with the current company established in 2018; engagements span frontier-model labs, voice-AI developers, AI compute vendors, computer-vision suppliers and autonomous-mobility programmes, with client identities withheld by agreement.
See AEO services and GEO services for scope, QA process for how review gates are defined, and the glossary for the terms used here.
Sources and further reading
- Aggarwal et al., "GEO: Generative Engine Optimization", ACM SIGKDD 2024 — across 10,000 queries: statistics up to +40% citation visibility, authoritative quotations roughly +30%, fluency +15–30%, keyword stuffing −10%.
- Google Search Central documentation on AI features and Search Console reporting — verify current behaviour before relying on any partner's description of what is reportable.
- Lifewood delivery figures (50+ languages, 40+ centres, 30+ countries, 56,788 contributors) are published on lifewood.com.