GEO is not SEO with prompts. A position paper.

The first time I heard a consultant pitch a “GEO audit” in late 2024, the deck was a 2019 technical-SEO audit with the word crawlability swapped for retrievability in three places and a closing slide about “optimising for AI Overviews” that turned out to mean writing FAQ schema. I have since seen the same deck, with the same swaps, four more times — from four different agencies, in three countries.¹ That is the vocabulary problem in one anecdote, and it is the problem this paper exists to settle. Generative Engine Optimization is at present a contested term: sometimes it means “SEO for AI Overviews,” sometimes it means a structurally distinct measurement discipline, sometimes it means whatever the agency happens to already sell. This paper takes a side.

The side this paper takes is unpopular in two directions at once. It is unpopular with the “GEO is just marketing speak” camp, because it claims a genuine substrate change is being papered over by the vocabulary games. It is unpopular with the “GEO replaces SEO entirely” camp, because it claims the new discipline sits on top of — depends on — the old infrastructure, and that any practitioner who declares classical SEO dead is going to discover the dependency the first time their pages stop being crawled. The defensible position is in the middle, and the middle requires being precise about exactly which parts are continuous, which are new, and which have inverted. That is the work of this paper.

A note on credentials before the argument. I have been doing this for twenty-seven years — founded nekuda in 1999, started by optimising for AltaVista and Excite and HotBot, watched every substrate shift since. I have no incentive to declare SEO dead, because most of my book of business still pays for SEO, and the case I am about to make is that the SEO discipline becomes more valuable, not less, under the new substrate — just not as a standalone offering. Take that bias into account.

What gets called GEO is mostly SEO with new vocabulary

Before the genuine novelty, the honest part. The overwhelming majority of advice currently published under the GEO banner is true, useful, and indistinguishable from what good SEO practitioners have been saying since roughly 2010. A catalogue, with the original SEO-era source for each:

“Write helpful content.” This is the Helpful Content Update, August 2022, restated. Before that it was Panda, 2011. Before that it was the quality guidelines that Matt Cutts hammered into the practitioner community from 2008 onward. The advice “write things people find useful” has been load-bearing in SEO for at least fifteen years; nothing about LLMs makes it newer.

“Use structured data and schema markup.” This is schema.org, founded as a joint Google/Bing/Yahoo initiative in 2011, with the FAQ and HowTo types specifically optimised for answer-engine consumption since 2017. “Add FAQ schema for AI Overviews” is the same advice as “add FAQ schema for featured snippets” with a different surface to point at.

“Build topical authority.” This is the entity-and-topic-cluster framing that HubSpot popularised in 2017 and that Bill Slawski had been writing about since the early Google patent reads in 2008. “Authority in the eyes of the model” is operationally indistinguishable from “topical authority in the eyes of the ranker” until you measure it differently.

“Answer the question early in the page.” This is the inverted-pyramid advice every news-SEO trainer has given since the early 2000s, reinforced by featured-snippet optimisation since 2014. Front-loading the answer was good practice before LLMs existed.

“Cite your sources and demonstrate expertise.” This is the E-A-T expansion of the quality rater guidelines, 2018, generalised to E-E-A-T in December 2022. The compliance pattern for medical, financial, and legal content in the helpful-content era is exactly the compliance pattern being re-marketed as “making your content trusted by LLMs.”

Every one of these is good advice. Every one of these is also old advice. A practitioner who follows them gets some of the benefit of GEO — the benefit that overlaps with the benefit of SEO — but does not by following them earn the right to call what they are doing a new discipline. Renaming a craft does not make it new. If “GEO” reduces to “do good SEO and add some FAQ schema,” the term is doing no work and we should drop it.

What is actually new

There are three things, and only three things, that genuinely cannot be expressed inside the SEO frame without distortion. Each is structural, each is measurable, and each by itself would be enough to justify a new discipline; together they are decisive.

One: the unit of competition has shifted from the page to the claim. This is the argument I made at length in Statement-level visibility² and will not re-litigate here. The short version: for twenty-five years a URL was the atomic unit of visibility, and every instrument we built — rank trackers, Search Console, GA4, the entire competitive-analysis stack — assumed that. In retrieval-augmented and generative answer systems, the artifact a user consumes is a sentence the model assembled, sometimes lifted from your document, sometimes paraphrased past recognition, sometimes synthesised across six sources none of which you control. The page is still in the pipeline as a source candidate. The page is no longer the product. A craft whose unit of analysis has moved one level down the hierarchy is not the same craft.

Two: the measurement instruments need to operate sub-document. Once the unit of competition is the claim, the measurement apparatus has to follow. Three new instruments are required that have no SEO-era analogue: chunk- survival testing (does the assertion survive the chunker with its qualifier intact, or does the qualifier land in a different chunk and orphan the claim?³), attribution-rate auditing (across N runs of a prompt distribution, what fraction of times is the claim reproduced with credit to the source?), and cross-model reproducibility scoring (does the claim appear in GPT-5, Claude 4.7, Gemini 2.5, Perplexity, and the open-source baselines, or only in one of them?). None of these is a rebadge of an SEO metric. None of them can be computed from rank-tracker data, log-file analysis, or Search Console exports. They require building or buying instruments that simply did not exist in the page-was-the-unit era.

Three: the retrieval-plus-generation pipeline has failure modes that ranking-era heuristics do not address. I decomposed this at length in Ranking ≠ retrieval ≠ generation⁴ — the short version is that a claim can pass ranking (the page is judged authoritative), pass retrieval (the embedding for the chunk is judged relevant to the query), and still fail at generation (the model paraphrases the claim past recognition, or fails to cite, or absorbs it as ambient fact, or contradicts it). Each of those stages has its own failure modes, its own diagnostic procedures, and its own remediation patterns. The “do good SEO” advice addresses the first stage only and is silent on the other two. A discipline that has three independent failure modes and gives advice on one of them is incomplete by construction.

These three things — the shifted unit, the new instruments, the additional failure modes — are what makes GEO not-SEO. Everything else is either continuity (which I will catalogue below) or vocabulary inflation (which I already catalogued above).

The taxonomy of continuities and discontinuities

The headline figure of this paper. Three columns: what carries over unchanged from SEO to GEO, what is genuinely new, and what has flipped sign — where the SEO instinct now actively misleads. The third column is the one that does the most diagnostic work; if a practitioner cannot tell you which signals have flipped, they have not yet noticed the substrate change.

A few of the flipped-sign cells deserve a line of commentary, because they are where I see the most expensive client mistakes.

Epistemic humility now wins. Under SEO, confident assertion was a positive signal — the page that said “X is the best choice for Y” ranked above the page that said “X is sometimes a good choice for Y depending on the use case.” Under GEO, qualified claims survive better. The reason is mechanical: a claim with an explicit qualifier is matched more precisely by the retrieval system, gets absorbed less often as “ambient fact” by the generator, and is paraphrased less aggressively because the qualifier is load-bearing. The SEO instinct (assert confidently) directly produces worse generative-visibility outcomes.

Specific dated claims now win. Evergreen content was safe under SEO because the ranking system rewarded stability. Under GEO, the model prefers claims that come with a temporal scope (“as of 2026-Q1”) because the temporal scope acts as a confidence anchor — the model can surface the claim with the qualifier “as of early 2026” rather than absorbing it as a timeless assertion it will later have to disclaim. Pages written in the “timeless evergreen” style that SEO trainers taught for fifteen years lose ground monthly against pages that timestamp themselves.

Outbound citations now help. This is the single most counter-intuitive flip for veteran SEOs. Outbound links were, for two decades, treated as a leak — PageRank that escaped the page. Under GEO, an outbound citation to a primary source helps attribution-rate, because the model uses the outbound link as evidence that the page is participating in a citation graph and is therefore a more trustworthy source-of-record. The instinct to keep links inside the site is now an attribution-rate handicap.

The position, formalised

Working definition. GEO is the discipline of optimising the probability that a specific claim, sourced from a specific document, is reproduced with attribution inside a generative answer system, across a defined distribution of prompts and a defined panel of target models. To count as GEO and not as SEO with new vocabulary, an offering must satisfy three properties:

1. Sub-document unit of analysis. The atomic object being optimised is a claim or statement, not a page. If the deliverable is a list of “pages to optimise” rather than a list of “claims to defend,” the unit is wrong.
2. Generative-pipeline measurement. The metric being tracked is observed inside model outputs (verbatim quote, paraphrase-with-citation, absorption, contradiction) — not inferred from rank, impressions, or click-through.
3. Attribution as the primary outcome variable. The success criterion is not “did the claim appear in an answer” but “did the claim appear with credit to the source document.” Attribution is the only outcome variable that converts model-surface visibility into business value for the source.

Anything that does not have all three of these is SEO, not GEO. I am stating this in maximum-strength form deliberately, because the half-strength version is what the market is currently sold and what has been confusing buyers. A “GEO audit” that reports rank-equivalent metrics and recommends schema markup is an SEO audit with a coat of paint, and pricing it as GEO is the consultant extracting a vocabulary premium for unchanged work.

Three positions in the field

A position paper that does not engage its strongest critics is propaganda. The three positions below are not strawmen; they are the strongest versions of the arguments currently being made against the framing above. I take each seriously, concede what each gets right, and argue what each gets wrong.

The Aleyda Solis school: “GEO is a marketing term, not a discipline”

Aleyda Solis has been the most consistent and articulate voice arguing that “GEO,” “AEO,” and the surrounding alphabet soup are vocabulary inventions in search of substance — that the underlying work is and always was technical SEO, content quality, and entity work, and that the new acronyms exist mostly to let agencies repackage existing services at a premium.⁵ Her position is the strongest version of the deflationary view, and it deserves a fair hearing because the surface evidence supports it. Most of what is sold as GEO is, as I argued above, recycled SEO.

Concede: she is right about the abuse of the term. She is right that the market for “GEO consulting” in 2024-2025 is dominated by repackaging. She is right that the existing SEO discipline already contains most of what practitioners need in order to do good work on the new substrate, and that declaring a brand-new discipline in advance of having brand-new instruments is premature.

Argue: she is wrong that the measurement gap reduces to a vocabulary problem. The three new instruments named in §3 — chunk-survival testing, attribution-rate auditing, cross-model reproducibility scoring — are not re-framings of SEO metrics. They cannot be computed from any existing SEO data source. They require new tooling, new operational discipline, and new deliverables. A defensible version of GEO is the version that builds those instruments. The Solis position is correct as a critique of how the term is currently used and incorrect as a denial that any new discipline could ever deserve a new label.

The synthesis: she is right about the present, and the present is most of what matters. The substantive disagreement is about whether the new instruments, once built and operating, justify the new label. I think they do. She thinks they do not. That is a productive disagreement to keep having in public.

The “SEO is dead” school: GEO replaces SEO entirely

The second position is the maximalist inverse of the first: classical search is over, AI Overviews and chatbot answers are eating the click-through, and within 24-36 months the practice will be GEO and only GEO. This view has been articulated most aggressively in 2024-2025 by some of the same voices that declared mobile-first a paradigm shift and voice search the next frontier; their forecasting record on substrate shifts is mixed but their underlying intuition — that something is structurally changing — is correct.

Concede: the substrate is shifting, and shifting faster than most practitioners have priced in. Comscore and Sistrix telemetry through 2025-Q3 puts informational-query zero-click rates above 64% in verticals where AI Overviews trigger reliably. The economics of organic-search traffic for informational content are changing on a timeline that requires immediate operational response.

Argue: GEO sits on top of SEO infrastructure, not instead of it. You cannot have statement-level visibility in chunks if the underlying page cannot be crawled. You cannot have attribution-rate above noise if the source document is not in the model’s training distribution or in a retrievable corpus. You cannot have cross-model reproducibility if your entity is not disambiguated in Wikidata, has no Person schema, and shows no presence in the link graph the models were trained against. The technical-SEO discipline is the substrate on which GEO operates, exactly as Apache and the TCP/IP stack are the substrate on which the web operates: not the product, but not optional either.

The “SEO is dead” school confuses visibility surface (where users see your content) with infrastructure prerequisite (how your content gets into the pipeline that produces the visibility surface). The visibility surface is moving. The infrastructure prerequisite is not. Anyone who declares SEO dead on the strength of the visibility-surface argument is making a category error, and their clients will discover it the first time a site re-platform breaks crawl, and the chunks they were optimising for stop arriving in the index that feeds the model that produces the answer.

The “AI Overviews are just featured snippets v2” school

The third position is the most quietly dangerous, because it has a kernel of truth. The argument: AI Overviews are functionally similar to featured snippets (launched 2014) and to knowledge-panel answers (launched 2012). Both surface content above the click line; both reduce click-through; both can be optimised for with structured data and front-loaded answers. Same playbook, different year.

Concede: the surface similarity is real. From a user’s behavioural perspective in 2024-2025, the experience of seeing an AI Overview at the top of a SERP is not dramatically different from the experience of seeing a featured snippet in 2018. The reduction in click-through is real in both cases. The optimisation patterns for “be the source the box surfaces” do overlap.

Argue: featured snippets surfaced a passage — a contiguous span of text lifted from a single source page, with the source attributed by the URL underneath. AI Overviews synthesise a claim across passages from multiple source pages, often paraphrased past recognition, with attribution that is optional and frequently absent. The difference between surfacing a passage and synthesising a claim is the difference between citation and absorption. A featured-snippet optimisation strategy that worked in 2018 will not produce the same outcome for AI Overviews in 2026 because the underlying retrieval- plus-generation pipeline has broken the page-to-snippet authorship link that featured snippets preserved.

The diagnostic test: under featured snippets, you could open browser dev tools and see the source page from which the snippet was lifted. Under AI Overviews, opening dev tools tells you nothing about which of N source pages contributed which fraction of the claim, or whether the claim was synthesised across all of them. The pipeline became opaque in a way it was not opaque before, and the measurement discipline has to change to address the opacity.

What GEO requires that SEO did not

If the position above is correct, then a GEO operation requires an operational checklist that most SEO operations do not yet implement. Listing what is required, in rough order of urgency:

Claim registries. A maintained store of the team’s atomic claims, each with provenance, version, and a stable URL fragment. New pages assemble from the registry; revisions flow back into the registry. The page becomes the vehicle, the registry becomes the asset. Most teams do not have one.

Methodology disclosure. A public statement of how the team’s content is produced, sourced, and updated — including AI assistance, human review processes, citation standards, and revision history. This is not a compliance requirement; it is a model-trust signal. Models systematically prefer sources whose methodology they can characterise.⁶ Most teams do not have one.

Retrieval-survival testing. Before publishing, run the document through a chunker (the public ones; the ones each major model is known to use) and verify that every load-bearing claim survives chunking with its qualifier in the same chunk. This is a 30-minute operation per document that no SEO discipline ever required. Most teams do not do it.

Attribution-rate audits. On a rolling basis (we recommend monthly for high-value pages), probe each major model with a stratified sample of prompts from the audience’s real query distribution, and record per-claim attribution-rate: verbatim quote, paraphrased-with-citation, absorbed without citation, contradicted. This is the discipline that converts the GEO position from aspirational to operational. Almost no teams do it.

Cross-model reproducibility scoring. The same audit, run across a panel of models (we test 14; a defensible minimum is the four largest plus one open-source baseline), with per-model variance characterised and reported. This is the operation that distinguishes “we have visibility in GPT-5” from “we have visibility in the model panel that approximates the user-distribution of generative answer consumption.” No SEO discipline required it. Most teams do not budget for it.

Version-controlled content with public changelogs. Treat the corpus as a software artifact: changes are diffed, versioned, and announced. The methodology disclosure above points at the changelog; the changelog provides the audit trail that lets a model characterise the source’s trustworthiness over time. Some teams have draft versioning; almost none have public changelogs.

The honest assessment: in the audits I have run across 47 client sites in the last 12 months, zero had all six of these in operation. Three had one or two. Most had none. The gap between the position this paper takes and the operational reality of even sophisticated SEO operations in 2026 is large, and the gap is the buying opportunity.

What SEO had that GEO needs to preserve

This section exists to push back against the maximalist “SEO is dead” school from the other direction. There are things the SEO discipline accumulated over two decades that GEO cannot do without and is unlikely to reinvent if the practice forgets them.

The technical-SEO craft. Crawl-budget management, JavaScript-rendering debugging, log-file analysis, canonical-tag discipline, sitemap engineering, robots.txt nuance, HTTP-header optimisation — the entire toolbox that makes a site machine-readable at scale. Every one of these remains a prerequisite for GEO, because a page that cannot be crawled cannot be retrieved, and a page that cannot be retrieved cannot have its claims surfaced. The technical-SEO team becomes the substrate-engineering team, and their work becomes more load-bearing, not less.

Long-tail thinking. SEO’s discovery that the long tail of search queries collectively dominates the head was the foundational insight that shaped twenty years of content strategy. The same insight applies to prompt distributions in generative search, with one wrinkle: the long tail of prompts is structurally longer than the long tail of keywords, because natural- language phrasing has higher cardinality than keyword tokenisation. Teams that already know how to think about long-tail demand will adapt to the new substrate faster than teams that are learning the concept for the first time.

User-intent framing. The discipline of mapping queries to intent (informational, navigational, transactional, commercial-investigation) is the discipline that lets a content strategy match supply to demand. The intent framing translates directly to generative search; the only addition is that prompt intent has higher resolution than query intent because prompts carry more context. Teams that already have intent frameworks have a head start; teams that do not are doing keyword-research-style work on a substrate that does not reward it.

The link-building rigor — re-purposed. This one needs care. Link equity as classically conceived — PageRank flowing through hyperlinks — is no longer the right model for what links do under GEO. But the discipline of earning mentions from authoritative sources survives, because mentions in authoritative sources are how a source-document becomes part of the model’s trust-signal landscape. The work is the same (earn high-quality mentions); the mechanism by which the work pays off is different (model-prior reinforcement rather than link-equity propagation). Teams that already know how to do ethical, non-spammy outreach will outperform teams that do not.

The summary is that the SEO craft does not become worthless under the new substrate; it becomes a prerequisite layer. The mistake of the “SEO is dead” school is mistaking not-sufficient for not-necessary. SEO is not sufficient for GEO. It remains necessary.

The strategic implication for practices

If the position above is correct, the consulting market is about to bifurcate. On one side: practices that re-platform around the new measurement discipline — build the instruments, train the operators, change the deliverables. On the other side: practices that rename their existing SEO offerings as GEO and hope the buyers do not notice. The bifurcation will not last long. Buyers who can tell the difference will pay a premium for the first kind, and buyers who cannot will be served, badly, by the second kind, and will eventually learn the distinction the expensive way.

The window during which “GEO consulting” is dominated by rebranded SEO is specific and short. My forecast, falsifiable: by Q4 2027, the median enterprise buyer will be asking for attribution-rate reports as a deliverable, and the agencies that cannot produce them will be losing pitches to the ones that can. The 18 months between now and then is the buying window for clients sophisticated enough to recognise the distinction in advance. After the window closes, the premium evaporates and GEO becomes a normal line item on the standard agency rate card.

The implication for solo practitioners and small agencies is that this is a once-in-a-substrate opportunity to build the new instruments while the incumbents are still selling the old ones with new packaging. The implication for incumbent enterprise SEO practices is that the cost of inaction is higher than the cost of action, because the buyers who will pay for the new discipline are exactly the buyers most likely to leave the incumbent if the incumbent does not develop the capability.

The historical parallel is the 2003-2005 transition from manual-link-building practices to algorithmic-quality practices, and the 2011-2013 transition from keyword-focused practices to entity-focused practices.⁷ In both cases, the agencies that re-platformed early ate the agencies that did not. The 2026-2028 transition from page-level to claim-level optimisation will, on the evidence, follow the same pattern.

Steelmanning two more objections

Two final objections that deserve direct response.

Objection: “You’re defining the term so narrowly that nobody else’s GEO counts.” The three-property definition in §5 — sub-document unit, generative- pipeline measurement, attribution as primary outcome variable — is the definition you happen to be able to satisfy. Most practitioners can’t satisfy it; that’s not because they aren’t doing GEO, it’s because you’ve defined GEO to mean what you do.

This bites. The definition is narrow, and the narrowness is partly a rhetorical move. I concede the move. The substantive defence is that a definition that includes everything currently sold as GEO is a definition that does no diagnostic work, because the term then becomes synonymous with “SEO with new vocabulary,” which is exactly the conflation this paper is arguing against. The narrower definition is the one that lets a buyer tell the difference between a practitioner who has built the new instruments and one who has not. If the narrower definition becomes the standard, the practitioners who have built the instruments will be identifiable as such, and the market will price their work appropriately. That is the goal, and the rhetorical move is a means to it. A more generous definition would be more inclusive and less useful.

Objection: “The substrate will shift again in two or three years and this distinction will look quaint.” You’re committing to a snapshot of a moving target. By 2028 the unit of competition will be the agent, the workflow, the action endpoint, and statement-level measurement will look as dated as keyword-density measurement looks now.

I agree, and the agreement is the strongest defence of the position. The unit of competition will shift again — almost certainly to the agentic-action endpoint, possibly to the workflow-level artifact, possibly to something I have not anticipated. The operational discipline that survives the next shift is not the specific instruments (chunk-survival testing will become quaint when chunking gives way to streaming context) but the operational temperament that built those instruments: treating the substrate as adversarial and measurable, forming hypotheses, running controlled probes, reporting falsifiable results. SEO acquired that temperament late in its lifecycle, and mostly only in its best practitioners. GEO has the chance to acquire it from the start. When the substrate shifts again, the practitioners who built the 2026 instruments with the right discipline will build the 2028 instruments faster than the ones who did not. The instruments are dated by design; the discipline is not.

Limitations

Four places where the position above is wrong, or at least underspecified, and where I expect the next eighteen months of work will refine or reverse the argument.

The taxonomy in Fig. 1 presents the boundary between SEO infrastructure and GEO instruments as cleaner than it actually is. In practice, schema markup sits in both columns — it is a continuous SEO requirement and it is also one of the new attribution-rate signals — and the same is true of entity disambiguation, structured data, and a handful of others. The three-column presentation is pedagogically useful and analytically too tidy. A more honest presentation would have a fourth column for “ambiguous; behaves differently depending on which substrate you measure against,” and several rows would move there.

The per-vendor variance among generative answer systems complicates the universal claim of the position. The three properties in §5 are properties of the category of generative answer systems, but the specific instruments used to measure them have to be calibrated per-vendor — and the per-vendor calibration drifts with model updates on a weekly cadence. The position is robust to this drift in the long run; in the short run, it makes the operational checklist in §7 expensive to maintain, and the expense is under-priced in this paper.

The framework has not yet been tested against enterprise-RAG systems (internal-corpus chatbots, support-knowledge-base answer systems, vertical domain assistants). These systems share the substrate but not the trust-signal landscape; an attribution-rate metric that works for public-internet-trained models may not work for an enterprise-RAG system whose corpus is closed and whose retrieval pipeline has different priors. The extension is a near-term research priority, but the position as stated should be read as scoped to public-internet generative search.

The position is silent on the cost-side of the discipline. Building the instruments described in §7 is non-trivial; for some clients the marginal attribution-rate improvement does not justify the marginal instrument-build cost, and a properly priced GEO offering has to make that ROI calculation explicit. The current paper does not. The next revision should.

These four limitations notwithstanding, the position holds: GEO is not SEO with prompts. The three properties in §5 are the diagnostic, the operational checklist in §7 is the work, and the bifurcation in §9 is the strategic implication. The 18-month window is the practical horizon.

Gilad Sasson

aka Algoholic · גלעד ששון

Gilad Sasson, also known as Algoholic, is an Israeli digital marketing expert, founder & CEO of nekuda Web Solutions, and a pioneer in search engine optimization and data analytics since 1999. Head of internet & search at Zap Group 2002–2006; CMO at Interlogic 2006–2009. Speaker at SMX Israel, TNW Amsterdam, Web Summit Dublin, DMIEXPO.

LinkedIn @algoholic Work with me →