How to get discovered in AI search

Full analysis

Two guys from Discovered Labs, Ben Moore and Liam Dunne, walked through how ChatGPT, Gemini, and Claude actually pick what to show you when you ask a question. The pitch: stop chasing citations, start getting your brand baked into the model's weights and into the sources it trusts. For anyone who buys or builds with AI, the practical question is whether any of this is real, checkable machinery you can act on, or a new SEO industry inventing a priesthood.

How hard is this to undo? Easy. Nothing here asks for a big spend or a lock-in. It's a content-and-measurement strategy question, reversible any month.

What's actually being decided: whether to fund a "get discovered inside AI answers" program, and if so, whether to trust the specific mechanics these guys describe. What sets the deadline: nothing hard. Google's WebMCP (a proposed standard for letting AI agents click buttons and fill forms on your site) is the only dated-ish thing, and it's early.

The Skeptic

Discovered Labs sells the cure and diagnoses the disease. Convenient. Ben Moore describes a five-step retrieval pipeline in confident detail, but nobody outside the labs can see it, so "roughly one-third of ChatGPT's slots go to Reddit" is an estimate from black-box probing, not a number you can audit. Then he admits most Reddit sources aren't cited. So the whole strategy rests on invisible influence you can't measure and he can't prove. The pricing trick is the one testable claim: publish directionally-correct pricing into "blank space" and get cited because the model won't hallucinate a number. That one I believe, because it's cheap to test yourself.

The Researcher

The bones are real, dressed up with borrowed names. Query fan-out (the model splits your question into several sub-searches), re-ranking, and consensus-checking across sources are genuine parts of how retrieval-augmented systems work. DeepMind's AGREE is a real paper on making models check answers against retrieved evidence. But "Claude moved to it when they moved to Fable 5.1" is asserted, not shown, and the claim that Reddit's human-feedback data leaves a traceable "imprint in the weights" you can influence is a big leap. You can't inspect a frontier model's weights from outside. Brand co-occurrence in training data probably matters. How much, and how you'd move it, is unmeasured.

The Compute Pragmatist

One claim deserves a hard look: weekly top-layer retraining, full backbone retrain every three months. If that were literally true across the frontier, it would reset the entire economics of these models, because a backbone retrain is a multi-hundred-million-dollar run. More likely Moore is blending real continual fine-tuning and index refreshes into one story. The retrieval index absolutely updates fast, daily even. The deep weights do not move weekly. That distinction matters for the buyer: your fresh content shows up in what the model retrieves quickly, but the idea that you're rewriting the model's brain by next Tuesday is oversold.

The Builder

Forget the theory. What ships Monday? Three things. Publish clear pricing and spec numbers where competitors leave gaps, then check in ChatGPT and Perplexity whether you get quoted. Get consistent brand messaging across independent sites so a consensus check finds agreement, not contradiction. And start structuring your forms and pages so an agent can complete them, because clients are already reporting agents booking demos. That last one is a real integration cost with a real payoff. The rest is content hygiene you're already paying for, pointed at a new reader.

The Open-Source Advocate

Notice what's missing from the whole conversation: open models. Everything here assumes a hosted black box you probe from the outside. If you run Llama or Qwen with your own retrieval setup, you don't optimize for the model's mystery weights, you just control the index and see exactly what got retrieved and why. The opacity Moore sells services against is a property of closed APIs, not of the technology. Enterprises worried about "will my brand survive to citation" have another path: own the pipeline.

Where they disagree

The Researcher and the Skeptic split on the weights story. Researcher says brand-in-training-data is plausible and probably real. Skeptic says an unmeasurable influence you can't prove is a sales tool, not a strategy. The Compute Pragmatist lands closer to the Skeptic on the specific claim: weekly backbone retraining is almost certainly not happening, and that claim is doing a lot of the "get in the weights now" urgency.

The Builder and the Open-Source Advocate split on where the leverage is. Builder wants to game the closed models everyone actually uses. Advocate says the whole game exists only because those models are closed.

What it hinges on

One belief: is influence you can't measure worth paying for? The pricing-into-blank-space trick is real, cheap, and testable, so run it this week. The consensus-alignment work is just good content discipline, do it anyway. The "get imprinted into the weights" program is the part to hold off on until someone shows a measurement with an error bar, because right now it's a story.

Before funding anything: ask any GEO vendor for a repeatable measurement of citation share with a stated margin of error. Moore himself says the tooling lacks statistical rigor. That admission is your test. If they can't bound their own numbers, you're buying vibes.

Prediction: Between now and Google's next major WebMCP milestone (a stable public spec or GA), no independent, reproducible study will show that a brand can measurably raise its influence inside a frontier LLM's weights (as opposed to its retrieved sources) by a stated amount with an error bar under 10%; the "get into the weights" pitch stays unfalsifiable through mid-2027.

Confidence: Medium. The mechanism is real but unmeasurable from outside a closed model.

Why: You cannot inspect a hosted frontier model's weights, so any claim about moving them is inferred from output changes that also move when the retrieval index updates, and there is no clean way to separate the two from the outside. Discovered Labs' own Ben Moore concedes the tooling lacks the statistical rigor to bound measurement noise, which means even the sellers can't produce the number. For the pitch to become falsifiable, a lab would have to expose weight-level attribution or someone would have to run a controlled study on open weights and prove it transfers to closed ones, and neither has a reason to happen by mid-2027. The retrieval-side claims, by contrast, are already checkable, which is exactly why those will keep working as marketing while the weights claim stays a story.

Revisit by 2027-06-30: We're right if no reproducible, error-bounded study demonstrates measurable brand influence on closed-model weights. We're wrong if any lab or independent researcher publishes a controlled result showing a named brand moved a frontier model's weight-level output by a stated, bounded amount.

The read-write agent shift is the part of this episode worth acting on before the theory settles. If clients are already reporting agents booking demos, structuring your site for machine-actionability has a payoff you can measure in conversions, not citations.