How to Get Your Brand Into AI Recommendations

The question in the title is asked in a slightly hopeful way, as though there is a form to fill in. There isn't. But the mechanism is less mysterious than the GEO industry makes it sound, and once you see it, the list of things you can do gets short and a bit boring.

What is happening when an engine recommends anybody

Someone types "best tools for X" into a grounded assistant. The assistant does not consult an internal ranking of brands. It runs one or more searches, retrieves a small number of pages, reads them, and writes a summary of what those pages collectively say. Then it lists the names that appeared, usually with the descriptions those pages used.

So being recommended is a chain with three links: a page that mentions you exists, that page is in the retrieval set for the question, and the mention is clear enough to survive being summarised. Break any link and you are not in the answer.

This is also why the answer changes when you re-run it. The retrieval moves. I have watched the same question return a different mix of names on consecutive days. Anybody promising you a stable position is describing a system that does not behave that way.

Step zero: find out where you actually stand

Everything below depends on knowing which pages supply your category's answers, and almost nobody has looked. It takes an afternoon by hand.

Write down fifteen to thirty questions a buyer would type. Full sentences, with the constraints in them, because that is how people talk to an assistant. "Best CRM" is a search query. "What CRM should a four-person agency use if we already run everything in Notion" is what actually gets typed, and it is the second kind that has an unclaimed answer.

Run each through a grounded engine. Perplexity is the easiest starting point because it cites openly on every query. For each answer, record four things: were you named, who was named instead, which URLs were cited, and which host each URL sat on. Then run the whole set again two days later, because the answers move and a single run will mislead you about what is stable.

Now count the hosts. The shape of that count tells you which lever below is binding. If five hosts supply half the citations, you have a gatekeeper problem and lever two is nearly the whole job. If the citations are scattered across forty hosts with no repeats, there is nothing to pitch and lever four is where your month should go.

One warning that cost me six weeks. Record whether each answer was grounded, meaning the engine actually searched. Ungrounded answers list brand names from memory with no sources behind them, and they look identical in the chat window. I once had 236 stored URLs that I thought were citations; 213 came from an ungrounded model and were just domains it remembered. Analysing that table would have told me confident nonsense.

Lever one: be readable

The floor. If an engine's crawler can't fetch your pages, or your homepage is a single line of marketing over a JavaScript bundle, you are excluded from direct retrieval and every mention of you has to be second-hand.

Concretely: let the AI crawlers in (GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot are the ones worth naming in robots.txt), serve real HTML for your important pages, put a plain sentence saying what your product is and who it is for somewhere a machine can find in the first paragraph, and keep your pricing public. Structured data and an llms.txt file are worth adding and are not worth agonising over. Our free readiness checker runs this set of checks; there is nothing magic in it and you can do most of it by hand.

The reason this is lever one and not the main event: it removes a reason to exclude you. It does not create a reason to include you. A large amount of GEO advice consists entirely of this lever, which is why so many people do all of it and see no change.

Lever two: be described by other people

This is the actual work, and it is where the ranked-list format earns its reputation. In the one sample we have published — 23 URLs cited by Perplexity across six buying questions in mid-2026 — seven were ranked lists and fourteen sat on a vendor's own domain. Ranked lists are a tiny share of the web and about a third of what got cited. They are dense with exactly what an answer needs: several named options, each with a short description and a use case.

So the highest-yield move is usually getting your name into the specific lists your category's engines retrieve. Not lists in general. The ones that get cited, which you find by running the questions and counting the hosts. The mechanics of landing them are in how to get added to a best-tools listicle.

The fourteen-vendor-pages finding matters too, and it surprised me. Half the cited vendor pages were ordinary blog posts and feature pages on small tools with no authority to speak of. Engines are willing to quote a small vendor describing itself. That makes publishing the page that answers a buying question directly a cheaper first move than most people assume.

Lever three: be described consistently

Models compress. When five pages describe you five different ways, the summary that comes out is vague, and vague names lose to specific ones in a list of options. When those five pages all say roughly "the cheap AI-visibility tracker for solo consultants", that phrase is what gets repeated.

This is a positioning problem wearing a technical hat. Pick the sentence, use it on your own site, put it in every pitch you send to a list owner, and use it in your own writing. Fight for it in the twelve words next to your name in someone else's article, which matters more than the link attached to it.

Lever four: be the answer to a specific question

The category-level question ("best X tools") is contested by everybody and usually returns the same four incumbents. The narrow question ("how do I see which listicles ChatGPT cites for a niche B2B category") has almost nobody answering it well, and the person asking it is much closer to buying.

Long-tail answer pages are the only part of this that resembles ordinary content marketing, with one difference: you are writing for retrieval rather than ranking, which means being blunt and answering in the first paragraph instead of building to it. Engines quote the paragraph that answers the question. If your answer arrives in paragraph nine after some throat-clearing about how the industry is changing, it gets skipped.

What we tried that did not work

We spent about a year on a theory that community threads are the main thing engines quote, and built a product around it. Two things went wrong with that.

The first is that it got us banned. Four Reddit accounts, permanently, over several months. Not for spam in the obvious sense; for being a new-ish account that kept mentioning a product in relevant threads, which is indistinguishable from spam from a moderator's side and is treated accordingly. There is no warning and no appeal that goes anywhere. Our product no longer asks anyone to post, because we could not find a version of "post helpfully about yourself" that survives contact with a moderation queue at any volume.

The second is that when I finally looked at the citation data, the premise was weak. One Reddit thread out of 23 grounded citations. Not zero, and not the dominant source the strategy assumed. The one that did appear was titled "Top 5 tools to monitor your brand's presence in AI", which is a listicle that happens to be hosted on Reddit, which arguably supports lever two rather than the community theory.

I am including this because the advice to go earn community mentions is everywhere, it is what I used to write, and the version of it that involves you doing the mentioning has a real cost that nobody attaches to it.

How long any of this takes

Grounded retrieval responds to new pages fast. A page that gets crawled can start appearing in answers within days, sometimes within a week. That is much quicker than the SEO timelines people bring to this, and it is the single best argument for doing the on-site work first: the feedback loop is short enough to learn from.

Getting mentioned somewhere you do not control runs on human time. A listicle editor takes one to six weeks to reply, if they reply, and then their update cycle is its own delay. Budget a quarter to see whether a pitching push worked.

Influencing what an ungrounded model recalls from training is not on any timeline you can plan around. It happens when the next model is trained, from a corpus you cannot audit, and treating it as a goal is how teams end up spending money on activity with no measurable endpoint.

The engines are not one channel

Worth splitting when you measure, because averaging them produces a number that describes nothing:

Perplexity grounds every query and cites five to ten sources. Easiest to influence, easiest to measure, smallest by usage. Most GEO tooling leans on it for exactly that reason, ours included, so be aware of the bias that introduces.

ChatGPT decides per query whether to search. The same question returns memory one day and citations the next depending on phrasing. It is also the one with the volume that matters commercially.

Google's AI Overviews sit on Google's index, so ordinary SEO carries over more here than anywhere else. It still is not a clean mapping: in one audit we ran, 58 of 96 cited pages did not appear in Google's own top 100 for the query that surfaced them.

The order, and what usually goes wrong

Readable, described, consistent, specific. In that order, but with a measurement step in front of all of it, because which lever is binding depends on your category. If five hosts supply half your category's answers, lever two is nearly everything and the rest is rounding. If citations are spread across forty hosts with no gatekeepers, lever four is where the wins are and pitching lists is a poor use of a month.

The common failure is doing lever one thoroughly, seeing nothing change, and concluding the whole thing is hype. Schema and crawler access and an llms.txt file cannot cause an engine to retrieve you. They only make sure that when something does cause it, you're legible.

If you want the longer version of the underlying mechanism, the GEO pillar guide covers it, and how AI engines choose brands to recommend goes deeper on corroboration. To see where you stand right now, the visibility checker answers "does it name me" and the citation source radar answers the more useful follow-up, which is who it is naming instead.