Generative engine optimization is not one practice: a per-engine investment framework

Executive summary

Four numbers anchor this paper, and none of them describe the same engine behaving consistently. The first says that when a buyer asks Claude a question, roughly two-thirds of the time it is answering from what it already knows rather than checking the live web, which means the lever that moves a Claude citation is not this week's publishing calendar, it is whether your brand made it into the model's training distribution in the first place. The second says that on Microsoft Copilot, more than a third of all citation share is already spoken for by three domains that are not going to be displaced by better content alone, which means the return on a content investment aimed at Copilot looks structurally different from the same investment aimed at an engine with a flatter citation curve. The third says the market itself is not settling into a stable hierarchy the way Google search did over two decades; it is still being actively fought over, at a pace that makes a program built around today's engine leaderboard obsolete within two quarters. The fourth, drawn from our own prior research, shows that even a single well-optimized page does not get treated equally by the five major engines, its citation rate spans a 35-point range depending entirely on which engine is asking.

Put together, those four numbers argue against the way most enterprises currently run generative engine optimization: as a single workstream reported against a single blended mention-rate number, with one checklist, roughly one budget split, and one cadence applied across ChatGPT, Claude, Copilot, Gemini, and Perplexity as if they were interchangeable surfaces competing for the same fixed pool of visibility. They are not interchangeable, and the framework below is built to replace that one-size assumption with a scoring model an enterprise can run against its own program this quarter.

Why one generative engine optimization checklist is a budget mistake

The instinct behind a single GEO checklist is reasonable on its face. Publish structured content, earn third-party corroboration, keep the technical plumbing clean, robots.txt, schema, crawlable HTML, and in principle every engine that reads the open web should reward the same signals in roughly the same proportion. That instinct held up reasonably well in 2024 and early 2025, when the practical differences between engines were mostly about scale, ChatGPT had more users, Perplexity indexed more aggressively, and the underlying selection logic looked similar enough that one checklist covering accessibility, structure, authority, and coverage, the four dimensions behind our enterprise GEO readiness framework, was a reasonable enough approximation of what mattered everywhere.

What has changed by mid-2026 is not that those four dimensions stopped mattering. It is that the engines applying them have diverged enough, in how often they retrieve, in how concentrated their citation pools already are, and in how fast their relative audience size is moving, that a single blended score now hides more than it reveals. A brand can be climbing on one engine and losing ground on another while its aggregate mention-rate dashboard reports a flat, reassuring line. That is not a hypothetical risk. It is the direct, measurable consequence of the three gaps this paper documents below, each backed by independent, named research published within weeks of this paper.

The clearest evidence that the underlying market will not hold still long enough to justify a fixed, undifferentiated program comes from Kevin Indig's "AI Halftime Report: H1 2026," published on Growth Memo on July 27, 2026. Between January and July, ChatGPT's share of AI-search activity fell from 78% to 56%, a 22-point drop in two quarters, while Gemini's share climbed to 30% and Claude's climbed to 10%, and Google AI Mode separately crossed 1 billion monthly active users. A shift of that size in six months has no real analogue in the twenty years of organic search rankings most SEO practice was built to survive. An enterprise that spent 2026's first half building its entire generative engine optimization program around ChatGPT's dominant position started the year with a defensible bet and is, by midyear, already several points behind a market that moved out from under it. Budget allocated evenly, or worse, allocated by inertia toward whichever engine had the most usage last year, is budget misallocated against a market this volatile. The three axes below are built to replace that inertia with a repeatable, re-scoreable model.

There is a second, quieter cost to the one-checklist approach beyond the share-volatility problem above, and it shows up in how enterprises measure their own programs rather than in how the engines behave. A blended mention-rate dashboard, tracking citations across all engines as one rolling average, structurally cannot distinguish between a program that is winning everywhere a little and a program that is winning big on one engine while quietly losing ground on another. Both scenarios can produce an identical, gently rising trend line. That is not a hypothetical failure mode; it is the direct arithmetic consequence of averaging five engines that, as the next three sections document, do not retrieve, cite, or hold audience share in anything close to the same way. A team reporting a single blended number to leadership is, in effect, reporting a number that cannot tell leadership which engine actually needs the next budget increment, which is precisely the decision a generative engine optimization program exists to inform.

None of this argues for abandoning a shared foundation across engines. The accessibility, structure, and authority signals behind our enterprise GEO readiness framework are still the right starting checklist for any engine, because a page an engine cannot crawl or parse fails everywhere equally, regardless of that engine's retrieval frequency, source concentration, or share trajectory. What changes once that baseline is in place is where the next incremental dollar goes, and that allocation decision is exactly what a single checklist has no mechanism to make. The three axes below are a layer on top of that foundation, not a replacement for it.

Axis 1: Retrieval frequency

The first axis asks a simple operational question with a non-obvious answer: when a buyer asks an engine a question relevant to your category, does the engine actually go check the live web, or does it answer from what it already learned during training, supplemented by whatever it has cached. The distinction matters enormously for where GEO effort should go. A high-retrieval-frequency engine rewards fresh, well-structured publishing on a short cycle, because it is genuinely re-checking the web each time it answers. A low-retrieval-frequency engine rewards something closer to durable authority: being well-represented enough, broadly enough, and consistently enough that you were already part of the corpus the model learned from, because a fresh page you shipped last week may simply never get checked.

Josh Blyskal's Profound research, "The state of AEO in 2026: Claude is not ChatGPT," published July 22, 2026, put a hard number on this for the engine most enterprises assume behaves like its competitors. Claude invoked live web search for only 36.6% of the prompts Blyskal's methodology tested. For the majority of tested queries, Claude answered from its trained or cached knowledge with no live retrieval step at all. And on the minority of prompts where Claude did search, 79.2% of its citations came from Brave's top 10 organic results, meaning even Claude's live-search behavior is heavily filtered through a single upstream search provider's ranking rather than an independent crawl-and-rank process of its own.

That combination makes Claude, on the current evidence, the clearest low-retrieval-frequency engine among the five this paper scores. It is not that Claude never checks the live web, 36.6% of prompts is not zero, it is that a majority of the answers a Claude user receives in your category are already decided before any crawler runs, by whatever Anthropic's training pipeline absorbed and by Brave's existing organic rankings on the minority of prompts that do trigger a search. A brand investing in a rapid publishing cadence aimed specifically at moving Claude citations within weeks is optimizing for a retrieval behavior Claude mostly does not exhibit. ChatGPT and Perplexity sit at the other end of this axis directionally, both engines have well-established live-browsing defaults built into their standard product behavior, and both were designed from the outset around retrieving and citing fresh web results rather than answering primarily from static training knowledge, though neither Blyskal's study nor any other source in this paper's research base publishes a directly comparable retrieval-frequency percentage for either, and we are not inventing one. The honest position is directional: ChatGPT and Perplexity behave, by product design and prior reporting, like higher-retrieval-frequency engines than Claude currently measures out to be, and that gap is real and actionable even without a matching hard number on the other side.

The measurement implication follows directly from the behavior. If a team is tracking week-over-week mention-rate movement on Claude the same way it tracks ChatGPT, and treating a flat week as a sign the content program stalled, it is very likely misreading its own dashboard. A flat week on a low-retrieval-frequency engine can simply mean the underlying training snapshot has not changed, not that the content failed to land. The more honest cadence for Claude specifically is a slower one, checked monthly rather than weekly, paired with a qualitative check on whether the brand's core facts, pricing structure, category framing, named differentiators, are being represented accurately anywhere Claude is likely to have absorbed them, since that is the lever actually available on this axis, not the publishing calendar itself.

Axis 2: Source concentration

The second axis asks how much of an engine's citation pool is already owned by a small number of mega-authority domains, versus spread across a long tail of sites a mid-sized enterprise could realistically break into. A low-concentration engine offers real headroom: a well-built page from a challenger brand can compete for citation share on close to even terms with an established competitor. A high-concentration engine is closer to a winner-take-most market, where the practical opportunity is not displacing the top few domains outright but earning placement on the surfaces they already control, through reviews, marketplace listings, or corroborating mentions rather than head-to-head content competition.

Ahrefs' research, "The 50 Most-Cited Websites in Copilot," using data through June 2026, is the clearest evidence of a high-concentration engine among the five this paper covers. Amazon alone accounts for 14.6% of Copilot's tracked citation share. Walmart accounts for 10.2%. Wikipedia accounts for 9.6%. Three domains, none of them a typical B2B publisher, together hold 34.4% of everything Copilot cites.

For a B2B enterprise, that table is less a direct competitive threat, most buyers are not asking Copilot to compare enterprise software against a Walmart listing, than it is a structural signal about how Copilot's citation logic behaves generally: it leans hard toward a small set of domains it already trusts at scale, which means the practical path to Copilot visibility for most enterprise categories runs through getting corroborated on the trusted third-party surfaces Copilot already favors, G2, Wikipedia itself where a category page exists, established review aggregators, rather than through a first-party content push competing directly for the same citation slots.

Contrast that against the concentration picture on the engines where we have our own data. Our anatomy-of-an-ai-citation research tracked mention rate for a single well-optimized page across four engines and found a 35-point spread: 74% on ChatGPT, 61% on Perplexity, 57% on Claude, 48% on AI Mode, and separately 39% on Gemini. That spread is not a concentration measurement in the same sense as the Copilot domain-share data, it measures how often one page gets cited rather than what share of all citations a handful of domains hold, but it establishes the same underlying point from a different angle: citation behavior is not evenly distributed across engines, and a page performing at 74% on one surface can be performing at roughly half that rate on another with no change to the page itself. Our 2026 AI citation study, run across 4,100 high-intent B2B queries, adds a further data point in the same direction: a source cited by ChatGPT was cited by Perplexity only 44% of the time, meaning even two engines with broadly similar live-retrieval behavior on Axis 1 still disagree on which sources deserve the citation once they do retrieve.

The same study's citation-rank data adds a further wrinkle worth naming: even on the two engines with the highest mention rates, average citation rank was not identical. ChatGPT placed a cited source at position 2.1 on average across the sample; Perplexity placed it at 2.6; Claude at 2.9; AI Mode at 3.4. A brand cited by all four engines equally often would still surface first, on average, on ChatGPT and last, on average, on AI Mode, which matters in practice because a citation buried at position three or four gets meaningfully less attention from a buyer skimming a generated answer than one sitting first. Source concentration is therefore not only a question of which domains get cited at all; it is also a question of where, within the citation list, a given engine tends to place a given source once it does decide to cite it.

The deeper reason those citation patterns diverge by engine, not just by topic, is that the engines are not running one shared trust model to begin with. Our reporting on how ChatGPT, Google, and Claude disagree on sources found OpenAI rejecting 99.39% of the Reddit pages it retrieves as candidates, Google selecting Reddit up to 60% of the time on the same underlying content, and Claude citing Reddit zero times across 139,601 sampled grounding sources. Source concentration and source trust are two sides of the same coin: an engine that concentrates heavily on a handful of domains, like Copilot, and an engine that applies a wildly different trust filter to the same platform, like the Reddit split above, are both telling you the same thing, that a single source-earning strategy built for one engine will not transfer cleanly to the next one.

Axis 3: Share stability

The third axis asks the question a budget conversation actually needs answered: if we win a citation or a mention-rate gain on this engine today, how much of that gain is still standing in two quarters. A high-stability engine behaves the way organic search rankings historically have, a placement earned this quarter is a reasonably safe bet to still matter next quarter, absent a specific algorithm change. A low-stability engine is one where the relative size and behavior of the audience itself is moving fast enough that the value of a placement can erode not because the content got worse or a competitor out-optimized it, but simply because fewer buyers are asking that engine the question at all.

Kevin Indig's H1 2026 halftime report is the sharpest evidence available that generative engine optimization as a category currently sits in a low-share-stability environment, full stop, across every engine it covers. ChatGPT's share of AI-search activity dropped 22 points, from 78% to 56%, between January and July 2026. Gemini's share rose to 30% over the same window, and Claude's rose to 10%. Separately, Google AI Mode crossed 1 billion monthly active users, moving it from a secondary surface into a default answer experience sitting inside the same search box Google has owned for two decades.

A 22-point swing for the category leader in six months has no real precedent in organic ranking volatility. It is closer to the kind of share movement you would expect from a genuinely young, unsettled market, which is exactly what H1 2026's data shows generative AI search still is. Rand Fishkin's SparkToro analysis, published June 9, 2026 and built on Similarweb clickstream data independent of Indig's methodology, corroborates the broader instability from a different angle: Google's zero-click rate reached 68.01% in early 2026, up from 60.45% in 2024, a nearly eight-point move in two years using an entirely separate data pipeline from Indig's engine-share tracking. Two independent analysts, using two different measurement approaches, landing on directionally the same story, a market where less is settled year over year than practitioners are used to, is stronger evidence than either finding alone. Neither Indig nor Fishkin is measuring exactly the same thing Blyskal or Ahrefs measured above; this axis is specifically about how fast the ground is moving, not about retrieval behavior or citation concentration, and it deserves to be scored on its own terms for that reason.

The reporting implication is the most immediately actionable part of this axis. Most enterprise GEO dashboards we encounter still run on the same annual or semi-annual review cadence teams inherited from organic SEO reporting, where a ranking shift of a few positions over a quarter was the normal pace of change. A 22-point engine-share swing inside a single half, the size of move Indig's report documents for the market leader, would be roughly two-thirds finished before an annual review cycle even caught it, and a program that only re-checks its engine weighting once a year is, in effect, choosing to run several months behind a market moving at this pace by default, not by exception.

The generative engine optimization scoring model

The table below scores each of the five major engines across all three axes, using the hardest number available for each cell. Where a hard, engine-specific figure exists in the research base above, we use it directly. Where it does not, we use a directional qualitative rating, High, Medium, or Low, rather than inventing a percentage to fill the cell. Readers should treat the quantitative and qualitative cells differently: the quantitative cells are measured findings from named, dated research; the qualitative cells are our own directional judgment, built from product design and prior reporting, and should be treated as a starting hypothesis to test against your own category's data rather than a benchmark with the same evidentiary weight as the hard numbers next to it.

Two patterns in that table are worth naming before the worked example applies it. First, no engine scores uniformly favorable or unfavorable across all three axes, which is the table's central finding as much as any individual cell. ChatGPT looks strong on retrieval frequency and source spread but is the least share-stable engine measured. Claude looks weak on retrieval frequency but is gaining share fastest in relative terms, off a smaller base. That mix is precisely why a single blended score across engines destroys the information a three-axis model preserves. Second, this table has real, honestly-labeled gaps, Copilot's retrieval frequency and both Copilot's and Perplexity's share-stability figures are not directly measured anywhere in the research this paper draws on. We are naming those gaps explicitly rather than filling them with an invented number, and any enterprise running this framework against its own program should treat those specific cells as open questions worth their own measurement, not settled inputs.

It is also worth stating plainly what this table does not attempt to do. It does not convert an engine's score into a projected traffic or pipeline number, and we are not going to manufacture that conversion here for the same reason we do not manufacture a retrieval-frequency percentage for Copilot above: no dataset currently ties per-engine share, retrieval behavior, and citation concentration to closed revenue with enough rigor to make that claim responsibly. What the table supports is a directional allocation argument, this engine deserves relatively more of the next dollar than that one, given how it retrieves, how concentrated its citations already are, and how fast its audience is moving, which is a materially different and more defensible claim than a revenue forecast built on three weeks-old external studies stitched together.

We run a version of this scoring exercise as the opening step on new generative engine optimization engagements, alongside the technical and content audits that establish where a brand currently stands on each axis before any budget gets reallocated.

A worked example

To make the model concrete, consider a hypothetical mid-market B2B SaaS vendor, illustrative rather than a specific named account, sitting down to plan next quarter's generative engine optimization budget using the scoring table above. The vendor currently splits its GEO effort roughly evenly across ChatGPT, Claude, Gemini, and Perplexity, with no dedicated Copilot workstream, a common starting allocation for a team that has not yet applied a per-engine model.

Applying Axis 1 first: because Claude scores Low on retrieval frequency, the vendor's plan to ship six new comparison pages this quarter and expect a fast Claude citation lift gets revised. Fresh publishing aimed at Claude specifically is deprioritized in favor of a slower-compounding play, making sure the brand's core positioning, pricing structure, and category framing are consistently and correctly represented everywhere Claude's training pipeline is likely to have absorbed them, third-party documentation, review sites, its own well-corroborated site content, since that is the material more likely to still be shaping Claude's answers whenever its next training cycle runs, rather than betting on a live-search citation that only fires on roughly a third of relevant prompts to begin with.

Applying Axis 2 next: the vendor notices it has no Copilot-specific workstream at all, and the scoring table's High source-concentration rating for Copilot argues against building one focused on first-party content competing for citation slots. Instead, the higher-leverage move is making sure the vendor's G2 listing, Wikipedia presence if a category page exists, and review-site profiles are current and complete, since those are the surface types a high-concentration engine already trusts, rather than authoring new Copilot-facing pages that would be competing, structurally, against Amazon- and Walmart-scale domain authority for the same citation slots.

Applying Axis 3 last: because ChatGPT scored Low on share stability, down 22 points in six months per the Indig data underlying this model, the vendor keeps its ChatGPT workstream funded, ChatGPT still holds the largest single share in the market, but shifts its measurement cadence from quarterly to monthly specifically for that engine, and earmarks a smaller, standing allocation for Gemini and AI Mode given both are gaining share in the same window. The Claude allocation, small in absolute dollars given its still-modest 10% share, is treated as a longer-horizon bet consistent with its Low retrieval-frequency score, not cut, but not expected to show a fast return either.

There is a fourth step worth naming even though the scoring table does not directly cover it: reconciling the plan above against Axis 3's instability. Because Gemini and AI Mode are both gaining share in the same window ChatGPT is losing it, the vendor's revised plan sets aside a standing, if currently modest, allocation for Gemini-specific structured content, treating its rising share as a leading indicator worth funding ahead of the data becoming undeniable, rather than waiting for Gemini to overtake ChatGPT outright before reallocating. That is the practical difference between a program built to react to share shifts after they show up in a lagging mention-rate dashboard and one built to anticipate them from the same H1 2026 data this paper's Axis 3 section already has in hand.

None of the dollar figures, page counts, or specific allocation percentages in this illustrative walkthrough are real client data; the exercise is meant to demonstrate how the three axes change a real allocation decision, not to report an actual engagement's numbers. The instructive part is not any single number in the example. It is that applying the same three-axis logic to a real program routinely produces a materially different budget split than an even, undifferentiated allocation across engines would, without requiring a bigger total budget, only a better-informed one.

What this means by role

No single role owns all three axes, which is the same structural reason our enterprise GEO readiness framework and enterprise AI agent readiness framework both work best as cross-functional exercises rather than a single team's checklist. The three-axis model above is meant to sit alongside those frameworks, not replace either one: readiness asks whether you can be cited or whether an agent can complete a task on your site at all, and this paper's model asks, once you clear that bar, where the next dollar of investment should go given how differently the engines behave once you are in the running.

What to do this quarter

The framework only earns its keep if it changes an actual budget line this quarter, not next year's planning cycle. Run the three-axis scoring exercise against your own program using the table above as a starting template, filling in your own engine-specific mention-rate and citation data wherever you have it in place of our directional ratings, since your own measured numbers should always outrank a directional placeholder once you have them.

This is the same discipline behind the work we have run for enterprise clients navigating exactly this kind of engine divergence. Our engagement with Zenity, a B2B agentic security platform, started from a citation profile built almost entirely around one engine's specific quirks, and the fix was not a bigger bet on that engine, it was rebuilding measurement so the program held up as the underlying share numbers moved. Our engagement with Arnica, a DevSecOps platform, needed a comparable rebalancing once its own per-engine data showed a workstream concentrated where the audience was, structurally, least likely to still be in six months. Neither engagement started from this exact three-axis model, since the research underneath it, Blyskal's Claude retrieval data, Ahrefs' Copilot concentration data, and Indig's H1 2026 share numbers, all published within the last several weeks, but both are proof of the same underlying principle: a generative engine optimization program built to survive the market moving, rather than to bet on it holding still, is the only kind worth running in mid-2026. Score the three axes, re-weight the budget, and re-run the exercise next quarter. The engines are not going to stop diverging on your timeline.

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.