Definition · 2026-07-14
What is the compounding layer?
A founder types the question her buyers ask into ChatGPT: best invoicing tool for small agencies. She has spent years earning the top organic spot on Google for that exact phrase. ChatGPT answers with a confident little shortlist. Her product is not on it.
Across town, a buyer does what buyers now do. He asks Perplexity which payroll providers handle contractors in Germany, reads the answer, opens a couple of tabs, and books a demo. He never sees a search results page. The vendors left out of the answer were never in the running, and none of them will ever know.
And in a third company, a marketing lead is trying to work out why they keep getting named. Gemini recommends them, warmly and often, for a use case they barely promote. She traces the answer back through its citations and lands on a comparison post from 2019, a stack of old review-site entries, and a Reddit thread nobody at the company has ever read.
Three rooms, one mechanism
These look like three different stories: an SEO problem, a buying-behavior shift, an attribution mystery. They are the same story. Between your content and your buyer there is now a layer, and the layer is a model. When someone asks ChatGPT, Claude, Perplexity, Gemini, or a Google AI Overview what to buy, the AI does not rank pages the way a search engine did. It composes an answer from everything it can retrieve about the category, and it names a handful of brands.
The model does not just read your website when it decides whether to recommend you. It reads what the world has accumulated about you. The founder in the first room optimized the page; the AI answer system was looking across reviews, comparisons, mentions, and category pages to decide which invoicing tools belonged in the answer. The buyer in the second room never saw Google's search results page at all. The company in the third room keeps getting recommended for a poor-fit use case because outdated comparison posts, review-site entries, and Reddit threads still associate it with that problem.
That accumulated record is the thing this site exists to study, so it needs a name.
Naming it
The compounding layer is the accumulated body of retrievable material associated with a brand, including citations, mentions, reviews, definitions, and original data, that determines whether AI engines such as ChatGPT, Claude, Perplexity, Gemini, and Google's AI Overviews recommend that brand.
It is called a layer because it now sits between your marketing and your buyer. When someone asks any of these engines what to buy, the answer is shaped by this wider public footprint. It is called compounding because it behaves like capital rather than like campaigns: the material accrues, references beget retrieval, retrieval begets further references, and the value builds on itself whether or not anyone is watching.
Two clarifications belong inside the definition, because leaving them out is how the term gets abused.
First, the compounding layer is not just the content you publish. The best available evidence points the other way: in Kevin Indig's Growth Memo analysis of what predicts AI brand mentions, the single strongest predictor was brand search volume, not any property of the brand's content. The layer is the whole footprint: how many people search for you by name, how often you are mentioned, reviewed, and cited, whether definitions and data trace back to you. Content and citations are the inputs you can most directly engineer, which is exactly why they get oversold. Anyone who tells you that publishing the right pages earns AI recommendations is selling you a mechanism the data does not support.
Second, the layer is owned, not rented. Rankings can be lost in an algorithm update and ad traffic stops the day the spend stops, but a citation in someone else's article, a review on a marketplace, a definition the models retrieve verbatim: these keep working. That is the compound-interest half of the name, and it is also the half that should make you patient. Compounding is famously slow at the start.
The part the skeptics get right
There is a live argument in marketing circles that GEO, generative engine optimization, is just SEO repackaged, or worse, brand marketing repackaged with a new invoice attached. The skeptics deserve a fair hearing, because they are substantially right.
The inputs to the compounding layer are the things good marketers have always built: demand for your name, reviews, press, mentions, useful original material. Nothing in the definition above requires new activities. If the entire GEO industry vanished tomorrow, the advice "be genuinely known in your category" would still cover most of it. And a good deal of what is currently sold as GEO is hype: tactics with no measurement behind them, tools that rank brands from single samples, promises about mechanisms nobody has isolated.
What has actually changed is the retrieval mechanics, and the change is not cosmetic. A search results page held ten links and a lot of ad slots; an AI answer names a few brands and moves on, so inclusion is closer to winner-take-most. The buyer often never clicks through, so you cannot see the loss in your analytics. And the answers themselves are unstable: SparkToro's research documents high inconsistency in which brands AI engines recommend for the same question, which means anyone quoting a single AI answer as your "AI ranking" is measuring noise.
So the honest position is narrower than either camp wants. The assets are old. The distribution layer sitting on top of them is new, opaque, and volatile. It can be observed, but not yet measured cleanly. People are tracking AI mentions, citations, and answer share, but there is no settled system for separating signal from variance, correlation from cause, or durable inclusion from a lucky answer on a lucky run.
What to do about it, which is mostly: measure
This site will not give you growth advice. Other properties do that; this one exists to establish facts. But the definition implies a short list of things a founder can act on without waiting for anyone's data.
- Ask the engines your buyers' questions yourself, several times each, before trusting any tool's screenshot of a single run.
- Audit your footprint the way an engine sees it: your name's search demand, your review coverage, who cites you and for what, not your on-page keywords.
- Treat anything you publish as an asset in the layer or don't publish it: content that compounds is not merely written to rank. It is useful enough to be cited, quoted, referenced, or reused after the first click.
- Distrust any vendor claiming a content trick earns AI recommendations. So far, the strongest signals appear to come from the broader brand footprint, not from a single page or format.
One question still lacks a clear answer: can a brand deliberately build its compounding layer and enter the AI answer layer on purpose? And if it can, how long does that take? Many people are trying to find that answer, including us. Starting this month, we are running an experiment: a fixed panel of buying questions against four engines on a schedule, variance reported, prompts published, with the first category snapshot to follow. The methodology page will show the full machinery.