For two decades the rule for inbound was straightforward. Rank on Google. Build pages. Earn links. Wait.
A meaningful share of B2B research traffic now happens somewhere else. Inside ChatGPT, Perplexity, Claude, Gemini, and the assistants embedded in operating systems and IDEs. The reader does not see your page. They see a synthesis of your page, alongside two or three others, written by a model that did not consult your analytics.
This shift has a name now. Generative Engine Optimisation, GEO, and the rules that produced top-three Google rankings will not produce top-three answers in an assistant.
What the model actually reads
Most marketing teams write for two readers. The user and Google. The new third reader is a model that compressed several tens of millions of pages into a representation of how concepts relate. It is excellent at extracting concrete claims with named entities and numbers. It is terrible at extracting hedged generalities.
A page that says "we help companies streamline their operations" disappears in an LLM. A page that says "our average customer reduces close-of-month from eleven days to four within the first quarter" has a chance. The first sentence has no entities, no numbers, no claim worth quoting. The second sentence is a citation waiting to happen.
This shifts the optimal copy style. The pages that get cited in assistant answers are the pages whose sentences can be lifted intact into someone else's writing. Concrete. Numbered. Sourced where possible. The slop tier of the internet is invisible to models now trained to refuse it.
The new shape of rank
In a Google result, position one gets roughly thirty percent of the clicks. Position ten gets three. In an assistant answer, there is no position one. There is a synthesis that either names you or doesn't. If you are named, the user often clicks through anyway because they want primary sources. If you are not named, they leave with the conclusion already formed.
The new dimension is not click-through rate. It is citation share. Of the assistants that answer "what is the best release management platform for indie labels", how many name you in the top three. Of those that name you, how many quote a phrase you wrote.
The unit of optimisation moved from the page to the sentence. The sentence that compounds is the one good enough to be plagiarised.
What to write differently
Three rules that have been earning citations across the studio's client work since late 2024.
- Lead with the claim, then prove it. Models reward extractive structure. A heading that says "Three reasons" followed by a numbered list with one concrete claim per item is many times more citable than the same content told as prose.
- Use the exact words your category uses, even when you'd prefer your own. Position yourself in the language already in the training set. A model cannot retrieve a phrasing nobody used to talk about you yet.
- Publish at least one piece per quarter that is a primary source. Original research. A small dataset. A benchmark you ran. Anything that produces a sentence other people will copy. Original sentences are the unit of citation. They are also, separately, the cheapest brand asset you can manufacture.
How to measure it
There is no Search Console for assistants yet. The current state of the art is unglamorous. Ask the four major assistants the same fifteen questions every two weeks. Score whether your brand appears, in which position, and whether you are quoted directly. Track the trend. Two quarters of consistent data is enough to know whether the content is working.
Profound and Mariana AI track this automatically now. Both are clunky. Both are still better than guessing.
Most teams will not do this and will assume the existing SEO playbook still works. They are about to discover that it doesn't, in the same window where their organic Google traffic falls thirty to fifty percent over the next four quarters as zero-click answers eat the search-results page.
The agencies who win in this window are the ones who shipped citable primary content first, in language the audience already used, and got named in the first generation of assistant answers that now drive an outsized share of consideration. The funnel did not break. It moved upstream into a layer you cannot pay your way into. You can only earn it by being worth quoting.