Search is being answered before it is clicked. A growing share of commercial questions now resolve inside an AI summary, a chat response, or an assistant that reads the web on the user’s behalf and returns a paragraph instead of ten links. For brands that spent a decade optimising for position, this is a different game with a different scoreboard. The question is no longer only “do we rank” but “are we the source the model reaches for, and does it represent us correctly.”
Marketing strategist Jaivin Karnani has spent the past two years building systems around this shift, and the pattern that keeps surfacing in that work is unglamorous: generative engine optimisation is mostly entity hygiene and retrievability, not clever prompting. The brands that get cited are the ones a machine can identify without ambiguity, verify against a second source, and quote without reformatting. Almost every failure traces back to one of those three.
Published: September 9, 2026. This article reflects current practice in generative engine optimisation and Karnani’s implementation pattern across consumer and services brands.
Ranking Optimises for Position. GEO Optimises for Being Quoted.
Classical SEO is a competition for a slot. Ten results exist, and the work is to occupy a better one. Generative retrieval is not a slot competition. A model assembling an answer pulls fragments from several sources, weighs them against each other, and synthesises. There is no tenth place. There is cited and not cited.
That changes what “good content” means in a specific way. A page engineered to hold a ranking often hedges: it covers every subtopic, buries the answer below a preamble, and avoids committing to a number or a position because commitment creates a surface to be wrong on. A page engineered to be quoted does the opposite. It states the answer plainly, early, in a sentence that survives being lifted out of context. Hedged prose ranks. Committed prose gets quoted.
The practical test Karnani applies is simple: take any paragraph on the page, remove it from the page, and read it alone. If it still makes a complete claim that a reader could act on, it is retrievable. If it only makes sense with the three paragraphs above it, a model will either skip it or paraphrase it badly.
The Entity Layer Is the Part Most Brands Skip
Before a model can cite a brand, it has to know what the brand is. That sounds trivial and is where most implementations fall apart. An entity is only useful to a machine if it is unambiguous, consistent, and connected.
Unambiguous means one canonical identifier. If a company appears as three slightly different legal names across its own site, its directory listings, and its press coverage, a retrieval system has three weak entities instead of one strong one. The fix is boring: pick the exact string, use it everywhere, and stop improvising.
Consistent means the structured data and the visible page agree. Structured data that claims a service area the page never mentions is a contradiction, and contradictions are exactly what verification layers are built to catch. Schema is not a place to make claims the page does not support.
Connected means the entity is linked to other entities that already exist in public knowledge graphs: a founder, a parent organisation, a location, a professional association, a body of published work. An isolated entity has nothing to be corroborated against. A connected one inherits confidence from its neighbours.
One failure mode deserves specific attention because it is common and self-inflicted: duplicated entity markup. Publishing the same entity block on many near-identical pages does not reinforce the entity. It fragments it. A retrieval system encountering fifteen copies of the same profile at fifteen URLs has to decide which one is canonical, and the usual resolution is to trust none of them strongly. One authoritative entity page, referenced by identifier from everywhere else, outperforms fifteen copies by a wide margin.
Machine-Readable Access Is Now Part of Technical SEO
Three technical surfaces determine whether a model can use a site at all, and all three are cheap to fix.
The first is crawler permission. Robots directives written for search engines are increasingly interpreted by AI crawlers with their own user agents. A brand that has quietly blocked those agents, often by inheriting a default it never reviewed, has opted out of generative visibility without deciding to. This is worth checking directly rather than trusting a dashboard, because monitoring tools misreport it often enough to matter. An empty Disallow directive, for instance, means allow everything, and tooling that treats it as a full block will report healthy sites as blocked.
The second is a machine-readable index of what the site actually offers. The emerging convention is a plain-text file at the site root that names the organisation, states what it does, and points to the canonical pages worth reading. It is not a ranking factor and should not be sold as one. It is a courtesy to a retrieval system that would otherwise have to infer structure from navigation markup, and a place to state the canonical entity name once, cleanly.
The third is rendering. If the substantive content of a page only exists after client-side JavaScript executes, some retrieval pipelines will see an empty document. The answer is not to abandon modern front-ends but to ensure the answer-bearing text is present in the server response.
How to Measure Something That Does Not Report Impressions
The uncomfortable part of generative optimisation is that the feedback loop is poor. There is no console that reports how often a brand was included in an AI summary. That absence tempts teams into two errors: declaring victory on activity, or abandoning the channel because it cannot be measured.
The workable middle is direct observation. Assemble the set of questions a qualified buyer would actually ask, run them against the assistants that matter, and record what comes back: whether the brand appears, which source is credited, and whether the characterisation is accurate. Repeat on a fixed schedule. It is manual, it is a sample rather than a census, and it is still far better than inference.
Karnani’s operating rule here is one he applies to every automated system: a check that cannot fail is not a check. A monitor that reports “healthy” whether or not the underlying job produced output is worse than no monitor, because it converts an unknown into a false confidence. If a GEO dashboard cannot distinguish “we were not cited” from “we did not measure,” it should say so.
The Correction Problem
There is a scenario every brand should plan for before it happens: the model answers confidently and gets it wrong. An outdated price, a discontinued product, a former executive, a merged entity described as independent.
Corrections do not propagate by complaint. They propagate by evidence. The reliable path is to publish the corrected fact clearly on the canonical page, ensure the structured data agrees, date it explicitly, and give the correction somewhere else to be corroborated: a profile, a listing, a piece of coverage. Then wait, because retrieval systems update on their own cadence. Brands that keep a current, dated, machine-readable source of truth recover from bad answers quickly. Brands whose most authoritative public statement is three years old do not.
An Implementation Sequence That Works
The order matters more than the individual items, because later steps depend on earlier ones being true.
Start by fixing the entity: one canonical name, one canonical profile page, one identifier referenced from everywhere else, and the removal of duplicate entity blocks. Then confirm access: verify crawler permissions against the live files rather than a report, and make sure answer-bearing content is server-rendered. Then align structured data with visible content so nothing is asserted that the page does not support. Then rewrite the highest-value pages so the answers are stated plainly and early, in paragraphs that survive extraction. Only then build the measurement loop, because measuring before the foundations are fixed produces a baseline of a system you are about to change.
Frequently Asked Questions
Is generative engine optimisation a replacement for SEO?
No. It is an extension. Nearly everything that makes a page retrievable by a model also makes it a better classical search result: clear structure, accurate markup, fast server-rendered content, and unambiguous identity. The difference is emphasis, not opposition.
Does publishing more content improve generative visibility?
Not on its own, and volume can actively hurt. Near-duplicate pages fragment an entity and dilute the signal that any single page is authoritative. One well-sourced page that states an answer plainly will out-perform twenty pages that circle it.
How long does it take to see a change?
Longer than a ranking change and less predictably. Retrieval systems refresh their underlying indexes on their own schedules, and some rely on snapshots taken well before a query is asked. Entity and access fixes should be treated as infrastructure with a multi-month horizon, not a campaign.
What is the single highest-return fix for most brands?
Consolidating a fragmented entity. Most organisations are describing themselves inconsistently across their own properties, and that inconsistency caps everything else. It costs almost nothing to fix and it is a precondition for the rest.
Closing Note
Generative engine optimisation attracts a lot of speculation because the mechanism is partly opaque and the vendors selling into it benefit from mystery. The practice underneath is far more mundane than the discourse: be identifiable, be accessible, be internally consistent, say the thing plainly, and measure by observation rather than by assumption. Those are not new disciplines. They are the same disciplines that have always separated sites that machines can use from sites that merely look finished, and the brands already doing them will find the transition to answer-first search considerably less disruptive than the brands that are waiting for a tool to do it for them.