Computational Marketing
FeaturesLong read

Gemini AI Overviews and the new engineering layer under GEO

Google's AI now selects passages over pages, rewarding claims that stand alone.

· 8 min read
Cover illustration for “Gemini AI Overviews and the new engineering layer under GEO”
Features · November 4, 2026 · 8 min read · 1,824 words

Gemini AI Overviews introduce a real engineering layer under GEO, but it is not a contest for the right positions within a page’s structure or token positions. The work is making information legible, attributable, and ready to synthesize. That distinction changes what teams build, how they distribute evidence, and what they measure.

Why Gemini selects content differently than Google's ranking algorithm did

Google ranking used to ask which page best matched a query. Gemini asks which information best explains a topic clearly and without risk. The unit of competition shifts from pages to passages. A document can cover a subject well overall while offering few claims that remain useful when separated from the surrounding text. Conversely, a passage can answer one part of a question without its page being the best destination for the whole query.

Query fan-out makes that distinction concrete. A single search becomes several sub-queries, each of which retrieves information that fits from wherever it lives. The resulting answer can therefore draw on several sources. A page participates as a collection of individually retrievable claims. Its usefulness depends partly on whether those claims can contribute to an explanation assembled from other material.

This also means that Google’s own AI surfaces need not select the same sources. Presenc AI measured low URL overlap between AI Mode and AI Overviews for the same query. That finding limits what a team can infer from success on either surface. Citation is not simply another name for ranking. Even within one company’s search products, the selection logic is not unified enough for visibility in one place to stand in for visibility elsewhere.

Why, then, has so much GEO advice focused on where an answer sits on a page? Layout is visible and relatively easy to change, while passage selection is harder to inspect. The resulting checklist offers concrete work in response to an uncertain mechanism. But that response can confuse the conditions that make content understandable with a formula that supposedly makes it win.

What the "technical GEO" checklist gets right, and where it stops being useful

Google’s granted patent “Layout-aware multimodal pretraining for multimodal document understanding” (US12393768B2) describes a model that breaks pages into blocks, including headings, paragraphs, images, and tables. It weights those blocks using position, spatial relationships, font, and style. Layout is therefore not an imaginary concern. A document’s visual organization can help a model distinguish its parts and understand how they relate. Treating a page as an undifferentiated string of text would miss that architectural capability.

Google’s Martin Splitt publicly introduced CenterpieceAnnotation and explained how page analysis breaks content into sections. The centerpiece section receives a higher weighting, while other sections receive less, and material on a different topic receives little consideration. This supports a sensible distinction between main content and peripheral content. Understanding which section is central is a different task from rewarding a fixed coordinate.

The checklist takes a further, less secure step. It turns these capabilities into prescriptions to place answers in the first viewport, pack entities into the first 2% of tokens, or write “3-sentence factual syntactic triads”. Some of these practices, such as server rendering and reducing client-side JS latency, are good defaults for how a site is built. Others give rules that look exact, but no scoring formula is shared. Neither widely cited AI Overview patent gives that formula, and descriptions of how the system is built do not show its source-selection weights.

Making an answer available early can help because it is easy to find and understand. Reducing JS delays helps when it makes substantive content available to retrieval. Those benefits survive without the coordinate theory. A useful engineering practice becomes misleading when its explanation promises control over a scoring system that the evidence does not disclose.

What synthesis-readiness actually requires at the structural level

The constructive response is to make independent legibility a default of the site’s architecture. For a technical founder, that means treating retrieval access and claim structure as build decisions. The page must offer information that can stand alone. A passage retrieved through fan-out should state what it asserts without depending on a teaser, a navigation label, or an unexplained hero headline to supply its meaning.

A static factual assertion at the top of each page serves that purpose. The opening statement establishes what the page is about in machine-readable form and gives retrieval something substantive to evaluate. That is a different design requirement from making every introduction short or forcing every answer into the same sentence pattern. The architecture creates space for a complete claim without pretending that completeness has a universal token count.

Availability is the next constraint on that claim. Client-side JS that defers substantive content creates a retrieval gap because Gemini cannot evaluate material it cannot read during the initial fetch. Server-side rendering closes that gap through how the site is built. Making the prose more precise cannot make text available for a content team. In this setting, rendering is part of the information contract between the site and the system trying to retrieve it.

Relationships need an equally deliberate structure. Schema tags declare that something is a thing, while a coherent entity graph expresses how things relate. Gemini’s entity and knowledge graph alignment tracks Google’s index, so the site’s internal organization helps shape how its topic is understood. Isolated labels are not the same as coherence. A site that describes connected entities consistently gives its individual claims a shared context, rather than leaving each page to establish that context from scratch.

A clear claim, readable delivery, and coherent relationships solve different parts of the same problem. None substitutes for the others. Even with a well-formed entity graph, content must be available to retrieval, and even with server rendering, an assertion must be precise. The engineering task is to make those requirements hold together by default.

Why off-site evidence matters as much as on-site structure

A legible site tells a model what a brand claims. It does not, by itself, establish that the claim is safe to repeat. Gemini’s preference for referenceable information depends partly on corroboration from independent sources. A brand publishing only on its own domain remains self-asserting.

Third-party pages account for a majority of brand mentions in AI answers, reflecting the importance of evidence outside the owned site. Digital PR contributes a meaningful share of LLM citations, yet only a small minority of GEO practitioners use it. Even a carefully built site can lack evidence from elsewhere that helps its claims fit into an answer drawn from several sources.

Stacker and Scrunch studied the distribution of the same content across third-party news properties and found a substantial median lift in brand citations across AI engines. The relevant mechanism is corroboration. But repetition and independent proof are not identical.

This is the architectural basis of the “phantom site” framing. It describes a content distribution strategy that puts material on third-party properties that models use as corroborating sources. The owned website is only part of the evidence footprint. Digital PR and third-party placements extend that footprint, while the substance being distributed determines what there is to cite. The strategy is not a substitute for having a claim worth supporting.

That substance comes from original survey data, proprietary case studies, and opinions from named experts whose conversations are documented. A model cannot generate those findings as original evidence, so it must source them. Distribution connects back to synthesis-readiness because a claim becomes more useful when its meaning is clear, its origin is identifiable, and supporting evidence exists beyond a self-description.

The measurement problem that makes GEO a guessing game without instrumentation

Google Search Console reports AI Overview impressions by page, country, device, and date, but it does not expose the question behind the answer, the answer text, or the competitors cited with a brand. That leaves a substantial interpretive gap. An impression can establish that a page appeared without explaining the selection context.

ChatGPT provides no official mentions data at all. Referral visits tagged through UTM parameters offer a noisy proxy. A visit and a mention answer different questions. The first records traffic reaching a site, while the second concerns whether a brand appeared in an AI response. Treating them as interchangeable obscures the very visibility that GEO is meant to measure.

Platform differences make that problem larger. ChatGPT and Perplexity show sharply different brand citation rates, and only a small fraction of domains are cited by both. A result on one platform cannot stand in for another. Model changes add another source of uncertainty, since a page that earned citations under Gemini 3.7 Flash may lose them under 3.8 Flash without a Search Console signal explaining the change. Citation selection can shift even when the team has not changed its page.

The practical metric is therefore visibility rate, which measures how often a brand appears in AI responses for a given prompt category, across multiple runs and multiple platforms. Repeated measurement addresses the variability that a single answer cannot reveal. Categorizing prompts also separates different kinds of success, rather than treating every mention as evidence that the brand is visible wherever potential buyers ask questions.

Internal data from a large set of accounts shows big differences between branded recognition and unbranded visibility across ChatGPT, Perplexity, and Claude. A model can know a brand when asked about it directly yet omit it from unbranded category queries. Recognition is not the same as category presence. Because those unbranded queries are commercially relevant, instrumentation needs to preserve that distinction.

Treating AI-overview legibility as a continuous engineering baseline, not a campaign

Core Web Vitals provides a useful parallel for organizing this work. Google introduced it as a ranking signal, and practitioners often treated it as a one-time audit. The teams that improved wired it into their CI/CD pipeline instead. The lesson for synthesis-readiness is to make the desired property part of how the site ships.

In practice, that means enforcing static factual assertions through page templates and treating server rendering as a build requirement. Entity consistency belongs in checks against the internal graph before publication. These are defaults. They move responsibility into the systems that produce pages, so legibility does not depend entirely on someone remembering a checklist for every release.

Off-site corroboration through digital PR and third-party placements belongs on a regular publishing cadence. Citation visibility across ChatGPT, Claude, Gemini, and Perplexity needs regular measurement and review. Traffic anomalies are too narrow a trigger. The monitoring loop must look at the responses in which the brand appears and check whether recognition leads to unbranded category visibility.

This approach gives teams a way to distinguish a site-level weakness from a change in models or selection logic and to keep improving what they control. A coordinate checklist fixes attention on one page at one moment. An engineering baseline makes clear claims, readable delivery, coherent entities, and attributable evidence part of every publication, while instrumentation reveals where those defaults still fall short.

More in Features