Does AEO work if nobody is searching for your category yet?

AEO works in a low-volume category when buyers are asking assistants about the problem, even if no keyword tool can see those questions. What stops it working is not thin search volume. Low volume is usually a measurement artefact, because a keyword tool counts short phrases and a buyer asking ChatGPT writes a sentence.

Our own query log is the cleanest example of that gap we can show you. Our Google Search Console and Bing Webmaster pull of 18 October 2026, described in our methodology, found that 80 distinct queries reached signalscite.com across 1,592 impressions in the preceding 90 days, and that 51% of those queries were six words or longer. The Bing query that ranks first and draws the most impressions for this site is a 38-word prompt about retrieval systems. None of those phrases carries a volume estimate anywhere, and all of them are real people asking real questions.

The thing that genuinely decides whether the work pays is whether the assistants are producing vendor answers in your category. Some categories get a list of five named suppliers. Others get a paragraph explaining the concept and a refusal to recommend. Those are different markets, and only the first one has a contest you can enter.

So the question to answer before spending anything is which of those two your category is in. That takes an afternoon and no budget, and the method is in the next section. A company that skips it is buying a retainer against a market nobody has confirmed exists, which is the single most expensive mistake available in this category.

How do you tell whether buyers are asking AI about your category at all?

You can tell whether buyers are asking AI about your category by running your buying questions as unbranded prompts and reading what comes back. Write down the ten questions your salespeople answer in every first call, phrased the way a buyer would say them out loud, then put each one to ChatGPT, Gemini, Perplexity and Google AI Mode without naming your company.

Three things in the answer tell you what you need. Whether the assistant recommends specific vendors or stays generic. Which companies it names if it does. Which pages it cites as sources, because that tells you what kind of page wins in your category, and it is usually not a vendor homepage.

Run each prompt more than once, on more than one day, before you conclude anything. Schulte, Bleeker and Kaufmann tracked four AI engines daily across four verticals for 45 to 46 days between 24 January and 20 March 2026 and found that the sources cited on consecutive days overlapped by only about 34% to 42%, with repeated runs inside a single 24 hour window overlapping about 32% to 43%. A single reading of a single prompt tells you almost nothing about your category, and a vendor who quotes you one is quoting noise.

Do the demand check before the site check, because it is the one that can stop the project. Google Search Console now carries a generative AI performance report covering AI Overviews and AI Mode, and per Google's documentation for it the report carries impressions only: no clicks, no position, no query breakdown, and no search type for those surfaces in the Search Analytics API. So it confirms that Google's AI surfaces showed your pages without telling you what happened next.

Why does a company nobody has heard of start further back than its pages suggest?

An unknown company starts further back because part of the retrieval decision happens before any page is considered. The engine partly decides what to search for from what it already knows, and a brand it has never encountered does not appear in that step.

That finding is the most uncomfortable one in the current literature for anyone selling page work, including us. Chen and Koudas, at the University of Toronto, replayed commercial questions through the engines' own APIs across two 90-day collections covering 158,099 answers and 468,037 citations, recording the sub-queries each engine wrote and the pages it then retrieved. ChatGPT wrote brands it already knew into its own sub-queries, preferred sites it knew, and in some cases named brands with no retrieved page behind them at all. Controlled experiments in the same paper showed brand names reordering product rankings when the underlying facts were identical.

The same paper puts a number on how much of this a page can touch. Retrieval results explained about 32% of the variation between pages in that study, against about 2.8% for page properties. Read that honestly and it cuts both ways: it is adverse to any claim that fixing a page predicts how often you get cited, and it supports the narrower claim that the one page-level property separating two pages on the same site is whether their titles and paragraphs match the questions people actually ask.

What follows is not that page work is pointless. It is that a prospect with no footprint is further from a citation than any page score can show, and that saying so is more useful than handing them a number. The work that closes the gap is partly off your own site, which our page on third-party mentions covers.

What should you do first if the category has demand but your brand has no footprint?

If the category has demand and your brand has no footprint, get into the retrieval path first and worry about everything else after, because the measured jump at that stage is larger than every stage after it. Being retrievable at all is the cheap half of the problem and the half you control.

Tannenbaum fitted a stage model to 34,960 unbranded prompt-engine observations across 2,854 monitored prompts between June and September 2026. Where neither the brand nor its own domain appeared in the observable retrieval path, the target brand was mentioned in 2.8% of GPT answers and 3.8% of Gemini answers. Where the company's own domain was cited, that rose to 49.0% and 58.4%. Worth naming the limits before anyone builds a plan on it: the study comes from a single commercial vendor, the same author published the cross-engine measurement cited elsewhere on this site, so the two are not independent sources, and the highest cell in that table has no misses in it rather than representing a law.

Practically, the order that follows from this is unglamorous. Make sure the crawlers can fetch and render your pages. Write a page for each buying question, worded as the question. Then go and earn the off-site presence that the assistants read, which in most B2B categories means the comparison pages, directories and communities they actually cite rather than your own blog.

Published research in your category is the version of this that compounds. Our study of how AI engines answer pharmaceutical process skid sourcing questions is a worked example from a sector where almost nobody had published anything checkable, and that absence is what made a single study worth citing.

Which situation are you actually in, and what does each one cost?

Four situations cover almost every company that asks this question, and the right move differs sharply between them. The table sets out how to tell them apart and what each one is worth spending on.

What the unbranded prompts do	What that means	What to do now	What it costs
Name five competitors, never you	Demand exists and the contest is live. You are absent from an answer buyers already see	Full programme: page coverage for the questions, then the off-site presence the engines cite	The highest of the four, and the only one with a competitor cost for waiting
Name you, but get your facts wrong	You are already in the retrieval path. The problem is what is said, not whether you appear	Correct the source pages the engines are reading, which is a narrower job	Small, and our AI answer accuracy page covers the mechanism
Explain the concept, recommend nobody	No vendor answer to win yet. The category is pre-retrieval	Publish the definitional and comparison pages that create the category, and re-test quarterly	Low. Content work, not a visibility programme
Cannot be reached or rendered at all	A technical failure upstream of every content question	Fix crawling and rendering first. Nothing else can be measured until that is clean	A fraction of a retainer, and often one afternoon of developer time

Most companies that ask us this question turn out to be in the first or the fourth row, and the two look identical from the inside. Our guide to checking whether AI can read your website separates them in about an hour.

Should you wait until your category has search volume before you start?

Waiting for search volume is the wrong trigger, because volume in classic search is a lagging indicator of behaviour that has already moved. By the time a phrase has measurable volume, the assistants have been answering the question for months and somebody is already being named in those answers.

The behaviour has moved further than most B2B marketing teams assume. 6sense's 2025 Buyer Experience Report, which surveyed nearly 4,000 B2B buyers, found 94% of them used large language models somewhere in the buying process, and that buyers first contacted vendors about 26.4 weeks into their journey. Most of the evaluation happens before anyone fills in a form, and an assistant is now part of that evaluation.

What that does not license is a claim that every category is equally affected. The 6sense figure is buyers using an assistant somewhere in the process, which is not the same as assistants producing a vendor shortlist for your specific question. The second is the thing that decides whether this work pays, and only your own prompt test can tell you.

So the trigger to watch is not volume and not the 94%. Re-run your ten prompts once a quarter and watch for the month the assistants start naming suppliers. That is the month the contest opens, and it is cheap to monitor while you wait. Working out whether to run the monitoring yourself is covered in agency versus tool versus in-house.

What else do people ask about AEO in a low-demand category?

The five questions below come up in assessment calls with companies that cannot find their category in a keyword tool.