A growing share of the questions your work could answer are now answered by an assistant instead. About 41% of people rely on AI summaries rather than clicking through, 13% skip search engines entirely, and among 18 to 24 year olds, 66% use ChatGPT to find information against 69% still using Google.
That raises a question with no settled answer: how do you end up being the source that gets summarised and named?
This page is shorter and less confident than the rest of the site, deliberately. The field has a name now, answer engine optimisation, and a consultancy industry has grown up around it. Far more is being sold about it than is actually known. What follows separates the parts with evidence behind them from the parts that are reasonable inference, and says which is which.
What is reasonably well established#
Being asked for by name still works#
The strongest finding available. When someone searches for a specific site or brand rather than a topic, AI Overviews are associated with an 18% increase in click-through. Topic questions get answered in place; name questions get sent onward.
So the most reliable thing you can do is be distinctive enough that people look for you specifically. That is a content and positioning problem rather than a technical one, and it is covered in getting found. Everything else on this page is marginal by comparison, and any service selling you the technical layer while ignoring this has the priorities backwards.
Being cited requires being readable#
If assistants cannot access your work, they cannot quote it. This matters more than it sounds, because blocking AI crawlers is a legitimate choice for licensing reasons and it removes you from this channel entirely.
That is a real trade-off with no obviously correct answer, and it should be a decision rather than a default. Licensing your work covers the other side, including how the RSL standard now lets you separate “may be used to answer questions about me” from “may be used as training data”, which is the distinction most working creators actually want and could not previously express.
Creator content shapes what assistants recommend#
Marketing analysts now treat creator material as part of the evidence base assistants draw on when recommending products, and brands have started commissioning creator content specifically to influence what AI systems say about them. Whatever you think of that as a practice, it indicates that this kind of material carries weight in the outputs.
Sources: Search Engine Journal, Fast Company, PwC on answer engine optimisation.
What is plausible but unproven#
These circulate widely in the AEO industry. They are sensible, consistent with how these systems appear to work, and nobody has published a controlled test. Treat them as reasonable guesses that cost little to follow.
Write in a way that can be lifted. Clear claims, one idea per paragraph, and the answer stated directly rather than buried after eight hundred words of preamble. If a model has to reconstruct your point from context, it is less likely to attribute it to you. The practical version: put the answer in the first two sentences under each heading, then explain.
Be specific and checkable. Numbers, dates, named sources, methodology. Material that can be verified appears to be favoured over material that cannot, which happens to be identical to what makes writing trustworthy to humans, so it is a low-risk bet either way.
Say something that is not already everywhere. If your page restates the consensus, there is no reason to cite you rather than any of the thousand other pages saying it. Original data, first-hand testing, your own figures and your own mistakes are the parts that can only come from you, and they are the parts worth citing.
Be discussed elsewhere. Assistants appear to weight what is said about you across the web rather than only what you say about yourself, so being quoted, linked and mentioned matters, which is another argument for the guest appearances covered in getting found.
Keep pages current and dated. Recency seems to help, and a visible date helps human readers judge you regardless.
Use structured data. Marking up articles, authors, organisations and FAQs with schema is long-established practice for search engines and costs nothing extra. Whether it influences assistant citation specifically is unproven, but it is free and it does no harm.
The llms.txt convention. A proposed plain-text file at your site root summarising your content for language models, by analogy with robots.txt. It exists, adoption is uncertain, and nobody has demonstrated that any major assistant reads it. Adding one takes ten minutes if you want to hedge. Do not let anyone charge you for it.
How to find out whether any of this is working#
You can measure this crudely yourself. A paid dashboard will also measure it crudely.
Write down ten questions someone would ask if they were looking for what you do. Ask each one in two or three assistants, from a logged-out or private session so your own history does not skew the answer. Record who gets named and linked.
Repeat monthly. What you are looking for is not a score but a direction: are you appearing at all, are you appearing for more of the ten over time, and who is appearing instead of you.
The important caveat: outputs vary between users, sessions and dates so a single check tells you very little and a trend over months tells you something. Anyone offering a precise “AI visibility score” is measuring a moving target and presenting it as a fixed one.
What to be sceptical about#
An industry grew up around this very quickly, and a good deal of what it sells is repackaged SEO with new labels.
Be wary of guaranteed citation in AI answers, because nobody controls that. Be wary of precise visibility metrics, for the reason above. Be wary of a technical checklist presented as the whole answer when the strongest available evidence points at distinctiveness and being searched for by name. And be wary of confident claims about how specific models rank sources, because the companies do not publish it and the people selling the service do not have access to it either.
If someone can show you a controlled result, a change made, a measurable effect, a method you could repeat, that is worth paying attention to. We have not seen one. If you have, we would like to.
The structural problem optimisation cannot solve#
Even done perfectly, this channel is worse for you than search used to be.
In search you got the visit. In an answer you get a citation at best, and most readers stop at the summary. That is where the traffic went. Assistants are a discovery channel where most of the value is retained by the assistant, and no amount of formatting changes that arithmetic.
So treat it as a supplement rather than a strategy. Be citable, because a citation is better than nothing and some people do follow through, and because being named in an answer is itself the thing that makes someone search for you afterwards. But build on channels you own, because that is where traffic goes when it goes anywhere at all.
What to do about it#
- Make something with a specific point of view that people would look for by name. This is the game; the rest is marginal.
- State your claims plainly and early, with figures and dates where you have them.
- Publish something only you could publish, your own numbers, your own testing, your own experience of getting it wrong.
- Decide consciously whether to allow AI crawlers, weighing this channel against your licensing position rather than defaulting either way.
- Add structured data and keep the dates visible. Cheap, harmless, probably helps.
- Run the ten-question check monthly and watch the direction.
- Do not spend money on AEO services until someone shows you a controlled result.
This is the fastest-moving subject on the site and the part most likely to be out of date or wrong. If you have measured results here, we would like to see them: [email protected].