Generative engine optimization has a timeline problem. Vendors sell it in quarters. The mechanics of how large language models actually learn what to say about a brand run on years. Duane Forrester's August 2 research explains why, and it does not contradict the other big AI-search finding making the rounds this year. It clarifies exactly what that finding is, and isn't, promising.
Two Studies, One Argument People Keep Flattening
Two pieces of research have been circulating in AI-search circles this year, and practitioners keep treating them as if they're arguing opposite things. They aren't arguing at all. Kevin Indig ran a 1,094-category study of ChatGPT and found that only 15.2% of categories have a stable, established AI-search owner — a brand that shows up as the consistent answer when someone asks the model who's best in that space. That leaves 89.3% of categories with no entrenched leader, which is the finding this site covered in our piece on category ownership durability: most of the AI-search map is still blank, and blank space is opportunity.
Then, on August 2, 2026, Duane Forrester published a piece on his Substack, Duane Forrester Decodes, that reads at first glance like the opposite argument. His claim: what a model says about a brand purely from its trained weights — no live web search, no retrieval, just what got baked into the parameters during training — is not something a content calendar can produce. He calls this 'parametric standing,' and his position is that it takes years to build because it depends on how many independent, differently-worded descriptions of a brand exist across the training corpus, written by people who don't work for that brand.
Read together instead of against each other, the two pieces describe different layers of the same system. Indig is measuring the competitive field of AI search as it exists today — who currently holds the mental shelf space in a given category, checkable by running prompts right now. Forrester is measuring the machinery underneath that field — how a brand gets onto that shelf in the first place, and how long the physics of model training make that take. One is a market-share snapshot. The other is a manufacturing timeline. Confusing them is how a genuinely useful piece of research, Indig's unowned-category finding, gets turned into a sales pitch for a genuinely unsupported promise: fast AI authority.
What Forrester Found: Parametric Standing Is Slow by Design
Forrester's argument isn't a hunch about how AI models behave. It's built on the actual composition of the data they were trained on, and the mechanics are worth sitting with because they explain why 'AI optimization' as a fast-turnaround service doesn't square with how these systems were built.
Start with the corpus itself. Forrester cites that the C4 training corpus — a foundational dataset used to train large language models — is dominated by content written between 2011 and 2019, with 92% of it falling in that window. Mainstream adoption of tools like ChatGPT is roughly a four-year phenomenon. That means most of what these models 'know' about a brand's reputation was written before most brands had any concept of AI search existing, let alone a strategy for it. You cannot retroactively write content into 2015. Whatever independent, third-party description of your brand exists from that era is largely fixed. Going forward, new mentions can add to the record, but they join a corpus where a single domain represents less than 0.05% of all documents — meaning even a brand's own site, publishing constantly, is a rounding error against the scale of everything else the model has ingested.
The second piece of Forrester's argument is about how models actually learn to recall a fact reliably. He points to research from Allen-Zhu and Li finding that a model needs 'sufficiently varied phrasing during pretraining' to extract a fact with confidence — meaning a claim about a brand has to be described many different ways, by many different, independent sources, before the model treats it as something it can state reliably rather than something it half-remembers. A single well-crafted case study on your own site, however good, is one phrasing from one source. It doesn't do the job varied third-party phrasing does.
The third piece closes the loop on why bigger models don't solve this for you. Forrester cites Mallen et al., whose research found that scaling up model size improves recall for already-popular, well-covered entities, but leaves long-tail or lesser-known entities roughly where they started. A newer, larger model isn't going to suddenly discover your brand's story. If your brand has thin, first-party-only coverage in the corpus, GPT-6 doesn't fix that any more than GPT-5 did. The fix has to happen upstream, in what gets written about you, by whom, before the next training run.
“The mechanism Forrester describes isn't a marketing problem you can outspend. It's a data-composition problem you can only out-wait, or start compounding against today so the payoff lands in the next training cycle instead of this one.”
What Indig Found: Most Categories Are Still Empty
None of that makes Indig's finding less true, or less useful. His 1,094-category study is answering a completely different, and much more immediately actionable, question: right now, today, does this category have a brand the AI model consistently names first? For 15.2% of categories, yes — there's an entrenched owner, and displacing them is a genuinely hard, multi-year fight, which lines up with everything Forrester describes about how parametric standing accumulates. For the other 89.3%, no. Nobody has locked it up yet. The model doesn't have a settled answer. That's not a training-data problem to wait out. That's a competitive gap to move into before someone else does the work first.
| KEVIN INDIG'S FINDING | DUANE FORRESTER'S FINDING | |
|---|---|---|
| What it measures | Whether a category already has an established AI-search owner | How fast a brand can build parametric standing from scratch |
| Time horizon | A snapshot — checkable by running prompts today | A structural constraint — measured in years, tied to training cycles |
| Core number | 89.3% of 1,094 categories have no stable owner | 92% of the C4 corpus predates mainstream LLM adoption |
| What it tells you | There's room to compete in almost every category | Winning that room through parametric standing alone takes years, not campaigns |
| Where it applies | Category-selection strategy and prioritization | Long-horizon brand-authority planning, distinct from retrieval-time citation |
The Reconciliation: Opportunity Isn't the Same as Speed
Put the two findings on the same timeline and the tension disappears. Indig tells you where the open ground is. Forrester tells you how long it actually takes to build a fortified position on that ground once you start. An unowned category is real opportunity — nobody's default answer, nobody's shelf space to knock off. But 'unowned' does not mean 'winnable by Thursday.' It means the field is open for a multi-year campaign to plant a flag that a future training run will actually pick up on.
This is exactly where the generative engine optimization pitch deck and the generative engine optimization mechanics part ways. Plenty of agencies are selling 90-day 'AI visibility' packages against Indig's 89.3% number, implying that because the category is open, it can be closed quickly with the right content sprint. Forrester's own research says the specific layer they're often selling against — parametric standing, what the model says unprompted from memory — doesn't move on a 90-day clock. It moves on a training-cycle clock, gated by how much independent, third-party, variably-phrased coverage accumulates between now and whenever the next major model gets trained on fresh web data. That's not a campaign. That's a multi-year content, PR, and topical authority in AI search program running continuously, with the honest expectation that its biggest gains show up a training cycle or two out, not in this quarter's report.
None of that means a 90-day engagement is worthless. It means it needs to be honest about which layer of AI search it's actually moving. There's a version of fast progress that's real, and it's not parametric standing — it's what happens at the moment of retrieval.
What Actually Moves Fast in Generative Engine Optimization
Forrester's whole argument is scoped to parametric standing specifically — what a model says with no live search running. Most consumer-facing AI answers today, in ChatGPT with browsing on, Perplexity, and Google's AI Overviews, involve retrieval: the model searches the live web at the moment of the query and pulls from what it finds. That's a different, much faster-moving mechanism. A page published this week can get crawled, indexed, and cited in a live AI answer within days, the same way it can rank in traditional search. That's the part of ai citation tracking that actually is fast, and it's worth treating as a distinct workstream rather than folding it into the same timeline as parametric standing.
That three-layer split matters for anyone building a real generative engine optimization program, and it's the difference between selling a category-ownership timeline honestly and selling it as a campaign. Retrieval-time work — structured content, clean markup, genuinely useful pages that answer the query a model is searching to fill — is closer to a standard content marketing and technical SEO motion, and it produces visible movement inside a quarter. Parametric-standing work is closer to a multi-year digital PR and third-party-coverage motion, closer to what a B2B SaaS or AI/ML company would run as sustained analyst, press, and community presence, not a content sprint. Both belong in the same program. Neither substitutes for the other, and selling one as if it were the other is the exact overpromise Forrester's research quietly indicts.
Do This Next: A Generative Engine Optimization Program Built on the Right Timeline
The practical move isn't picking a side between Forrester and Indig. It's building a program that runs both clocks at once, honestly labeled.
One more thing worth saying plainly to anyone evaluating a generative engine optimization proposal this quarter, including one weighing link building and third-party coverage work as part of the parametric-standing bet: ask which of the three layers above the proposal is actually promising to move, and on what timeline. If a vendor points at Indig's 89.3% number to promise fast category ownership without ever naming the parametric-standing mechanism Forrester describes, that's the gap to press on before signing anything. The category being open is real. The clock for owning it is longer than a campaign, and the only honest programs are the ones that say so while still moving fast on the part of the system that genuinely can move fast.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Tyler leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.