A prospect last month opened the call by telling me he had already shortlisted three agencies. I asked where he had found us. He had not run a single Google search. He had asked an assistant what to look for when hiring a paid media agency, asked a follow-up about what it should cost, and taken the three names that came back.
That is a different funnel from the one most marketing plans are built around. There was no ranking position involved, no click from a results page, no impression we could see anywhere. By the time he reached our site, the shortlist was already made and the site's only job was to not contradict what he had been told.
This is worth understanding properly, because there is already a great deal of nonsense being sold about it.
This is citation, not ranking
Search gives you a position in a list and a click. An assistant gives an answer, and sometimes names sources inside it. There is no second place: either you are in the answer or you are not.
That changes what the work looks like. You cannot optimise for a position that does not exist. What you can do is make it easy for a machine to extract a correct, specific, quotable statement about what you do — and make sure those statements exist in places the machine actually reads.
The second half of that sentence is the part people skip, and it is most of the job.
The two kinds of crawler, which nobody separates
If you look at your server logs you will see bots from OpenAI, Anthropic, Google, Perplexity and others. They are not all doing the same thing, and the difference decides whether any of this matters this quarter or in two years.
Training crawlers read pages to help build future models. GPTBot, ClaudeBot and Google-Extended are in this group. Being read by them might mean a model knows about you after its next training run. That is slow, indirect, and nothing you can schedule around.
Retrieval crawlers fetch pages live, during a conversation, when somebody has just asked a question. OAI-SearchBot, ChatGPT-User, Claude-User, Claude-SearchBot and PerplexityBot sit here. This group decides whether you can be cited today.
Plenty of sites block all of them, usually by pasting a disallow list someone published without reading what each agent does. The result is a site that ranks perfectly well on Google and is invisible the moment a buyer asks an assistant instead. If you want to be found this way, check your robots.txt and make sure the retrieval agents in particular are allowed.
What actually gets cited
Having watched which of our pages get pulled into answers and which do not, the pattern is fairly consistent.
Content that answers a question in one self-contained passage. Assistants extract chunks. A paragraph that only makes sense after three paragraphs of build-up rarely survives the extraction. Write the answer first, then the reasoning.
Specific, checkable statements. "We deliver outstanding results" is unquotable — there is nothing in it. "Retainers start at $500 a month and the ad budget is paid directly to the platform" is a fact a machine can repeat without risk. Published prices are unusually valuable here, because most competitors hide theirs and an assistant asked about cost has very little to work with.
Stated limitations. This is the one that surprises people. Saying who you are not for, and what you will not promise, makes everything else safer to quote. A source that admits a boundary reads as more reliable than one that claims to suit everyone, and models are reasonably good at picking up on that.
Structure a machine can parse. Headings that match real questions, FAQ sections, plain HTML text rather than text baked into images. Structured data helps at the margin; readable prose helps more.
What does not work
Keyword stuffing, in a new costume. Repeating "best performance marketing agency" forty times does nothing. These systems are not matching strings.
Writing pages addressed to the model. Anything that reads as instructions to an AI rather than information for a person is treated as what it is.
Volume. Publishing fifty thin articles to "cover more queries" is the same scaled-content pattern that earns a manual action in ordinary search, and it will cost you the organic traffic you already have. I have refused to do this for clients and would refuse again.
Anyone selling a guarantee. Nobody can make an assistant recommend you. If that is being promised, you are being sold something that cannot be delivered.
The uncomfortable part: most of it is not on your site
Assistants synthesise from across the web. When somebody asks for agency recommendations, the answer is usually assembled from directories, review platforms, forum threads and roundup articles — not from the agencies' own websites, which every model correctly treats as self-interested.
So the highest-leverage work here is the least technical. Being present and reviewed on the platforms your category uses. Being mentioned in comparison articles. Having real people say real things about you somewhere that is not your own domain.
That is slow and it is largely sales and relationship work rather than marketing work. It is also why a brand-new site with no mentions anywhere will not get cited no matter how well its markup is arranged.
You will struggle to measure it
There is no rank tracker for this. Assistant referrals mostly arrive as direct traffic, or with a referrer you will not recognise, and a large share of the influence leaves no trace at all because the buyer never clicked anything.
Two things that do work. Filter your server logs for the retrieval agents to confirm you are being fetched at all. And ask every inbound lead how they found you — then actually record the answer. When people start saying "I asked ChatGPT", you will know before any dashboard tells you.
A reasonable order to do this in
- Check robots.txt and stop blocking the retrieval crawlers.
- Publish the facts a buyer needs: what you do, what it costs, who it suits, who it does not.
- Rewrite your main pages so each section answers one question in one passage.
- Get listed and reviewed on the platforms your category uses.
- Ask leads where they came from, and keep a tally.
None of that is exotic. Most of it is the same discipline that made a site good for human readers, applied with the knowledge that the first reader is now often a machine summarising on someone's behalf.
The short version
You cannot rank in an answer, only be cited in one. Citations go to sources that state specific, checkable things, admit their limits, and are corroborated somewhere other than their own website. The technical part takes an afternoon. The reputation part takes a year, which is why starting it now matters more than the markup.
Two related reads: commercial-intent pages beat informational ones for the same reason answerable passages beat padded ones, and a brand that reads like every competitor gives an assistant nothing distinctive to repeat about you.
This is part of how we approach SEO now. If you want to know whether assistants can currently reach and quote your site at all, the free growth audit includes that check alongside the ad account review.