Intelligence
This page is the briefing we give every client in the first hour. It is free, it is sourced, and if you read it carefully you will be able to audit our proposal better than most agencies would like.
The mechanism
Influencing the model's own weights is extremely hard and extremely slow. Almost everything that works in practice touches the retrieval layer instead.Ethan SmithCEO, Graphite — Lenny's Podcast. Paraphrased from his description of the two-layer system.
There are two layers inside every answer engine, and confusing them is the single most expensive mistake in this field. The core model was trained months or years ago; changing what it knows is close to impossible and, where possible, takes about a year to show up. The retrieval layer runs a live search at the moment you ask, and rewrites the answer from whatever it finds.
Practically every tactic that has ever been shown to work in AEO acts on the second layer. Which means the common objection — surely we just wait for the next training run — has the mechanism backwards. You are not waiting for anything. The search is happening now.
Why rankings stopped deciding
In search, one link wins the click. In synthesis, the model reads across everything it retrieved and writes a single paragraph. Repetition across independent sources is what survives that compression.
ChatGPT's citations overlapped Google's results roughly 35% of the time. Perplexity's, closer to 70%. Ranking well on Google buys you a third of the way in on one platform and most of the way on another. Graphite citation study across thousands of questions.
Listicle and comparison content was the single most-cited format across a database of 177 million citations — blog and opinion content came second at roughly 9.9%. Profound internal citation dataset, reported by their AI strategist.
Relevant links tied tightly to your subject reportedly outperform chasing a thousand generic ones. The unit of value moved from volume to topical proximity. Mike King, iPullRank. Agency observation, not a published study.
As long as the engine only makes one or two backend search calls, it is structurally incentivised to grab one already-aggregated page covering every competitor at once.Josh, AI strategist at ProfoundParaphrased — his explanation of why listicles currently dominate, and why he expects deep-research and fanout modes to erode that advantage.
Where this is heading
Anyone selling you a permanent playbook is selling a snapshot. Here is what has already shifted, and what practitioners with data are expecting next.
Already happened
AI overviews now appear in close to six in ten Google searches, and when one appears, click-through to the underlying site roughly halves. The zero-click problem stopped being a prediction.
Already happened
Around 25 words in a chat interface against roughly six in a search box. The long tail did not shrink — it became the majority of the surface.
Happening now
Roughly 40% of answer-engine queries classify as purely generative — asking the model to build the spreadsheet rather than find the article about spreadsheets. The classic intent categories are being squeezed into a shrinking share.
Next
Shopping cards, maps and booking flows are already appearing. The expectation among practitioners is that answer engines follow the same path Google's results page did, adding vertical-specific clickable surfaces one at a time.
Next
Deep-research and fanout modes let an engine visit individual product and review pages instead of leaning on one pre-aggregated page. When that becomes default, the citation mix redistributes.
Next
Native transactions inside the assistant would move the measurement problem again — from no click at all, to a click that never touches your analytics.
Sources: AI overview figures from Asia Forest, HubSpot. Query length and clickability predictions from Ethan Smith, Graphite. Generative query share and listicle mechanism from Profound. Predictions are labelled as such.
The question everyone asks
Two different claims usually get argued as one. They need different evidence, and they have different answers.
Claim one: people are abandoning search. The available usage data does not support this. LLM visit volume has been measured at roughly one twenty-fifth the size of search, growing on top of Google volume that is flat to slightly down — a few percentage points, not a collapse. Total search activity may be roughly doubling once you count the queries agents run on a user's behalf, since a single session can trigger several.
Claim two: your specific category is being reshaped. This one is often true, and it is the one worth acting on. Impact varies sharply by vertical, and for particular use cases a single prompt genuinely replaces what used to be five searches.
Estimated share of traffic influenced by AI search, depending on category — with more tech-forward companies at the top of that range and 5–10% described as more typical. Graphite co-founder. Estimate, not measured across a defined sample.
One honest caveat carried over from the source: the practitioner making the macro argument states plainly that he does not have data proving total conversions are rising across the market — only that usage is additive. We repeat the caveat because it is the part most people drop.
Surfaces
Which is why a single "AI visibility" number, averaged across platforms, hides more than it tells you. We track them separately for this reason.
| Surface | Overlap with Google results | Behaviour worth knowing |
|---|---|---|
| ChatGPT | ~35% | Largest volume. Leans on a small number of backend search calls, which is part of why aggregated comparison pages do so well. Does not render JavaScript. |
| Perplexity | ~70% | Far smaller volume, but reported click-through roughly six to ten times higher than ChatGPT. Shoppable and sponsored answer formats already live. Does not render JavaScript. |
| Google AI overviews | By definition high | Appear in close to 60% of searches; click-through to sites roughly halves when present. Google does render JavaScript, which is why a page can work here and fail elsewhere. |
| Gemini & others | Varies | Consolidation between overviews, AI mode and Gemini is widely expected. Worth tracking separately until that settles. |
Overlap figures: Graphite citation study. Click-through comparison and shoppable formats: Profound. AI overview prevalence and click-through effect: Asia Forest, HubSpot. JavaScript rendering: Mike King, iPullRank.
A worked example
A field this young manufactures consensus by repetition alone.
The clearest case study is llms.txt — a proposed file that a great many people now believe answer engines read. Trace it back and the picture thins considerably: the idea was floated informally, repeated widely, and the repetition itself became the evidence. A Google search advocate has said publicly he was not aware of it being used in practice. Ask ChatGPT directly and it reports that it does not use one. By one account, even the site of the person credited with proposing it does not implement it.
Site owners then read unrelated evidence — their site getting crawled at all, which happens regardless — as confirmation the protocol works. That is the entire mechanism by which a large share of AEO "best practice" got established.
We are not immune to it either. Which is why everything on this site carries a source and a confidence tier, and why we run a control group before we tell you something works.
The citation report
This page gets updated as the field moves. The newsletter is how you find out when it did.