Getting Cited by AI Answers: What Is Known and What Is Guesswork
AEO and GEO are being sold with confidence nobody has earned. Here is the small amount that is actually established.
CTFM Team
There is now a service category selling optimisation for AI answers. Answer engine optimisation, generative engine optimisation, LLM SEO, pick your acronym. The pitches are confident and the case studies are glossy.
The honest position is that almost nobody has replicated, inspectable evidence about what makes a model cite one source over another. That does not mean do nothing. It means be careful about who you pay.
Search increasingly answers without sending a click. We covered the strongest available figure and its caveats in most Google searches no longer send a click. Whatever the exact number where you sell, the direction is not seriously contested.
Traffic arriving from AI assistants is currently small. As a share of total sessions, referrals from chat assistants remain a low single-digit percentage in most published analyses. It is growing, and it is not yet where your visitors come from.
Everything after those two points gets shakier fast.
There is a popular claim that visitors arriving from AI assistants convert far better than visitors from ordinary search. It may well be true. Look at the published numbers and you find:
Claimed lift over organic search
Roughly what is claimed
About 30% higher
One analysis of ecommerce sites
Around 4 to 5 times
A search-tool vendor's aggregate data
Around 23 times
Another vendor, based on its own signups
These are not small disagreements. They span nearly two orders of magnitude, which tells you the underlying measurements are not comparable: different industries, different definitions of a conversion, different attribution windows, and in most cases a company with a product to sell doing the measuring on its own data.
How to read any AI-search statistic
Ask three questions. Who measured it, and do they sell something that gets easier to sell if the number is big? What counted as a conversion, since a newsletter signup and a purchase are not the same event? Was traffic self-reported or attributed, given that assistant referrals are notoriously hard to track cleanly? Almost every striking figure in this space fails at least one of these, which is why they contradict each other so violently.
The underlying idea is still plausible for a mundane reason: someone who arrives after asking an assistant a specific question has already been filtered. Plausible is not measured, and you should not budget against it as though it were.
Here is the useful part. Almost everything credible people recommend for AI visibility is something you would do regardless, which means the downside of doing it is close to zero.
Answer the question directly, near the top
State the answer plainly, then explain. This helps a human skimming your page and it happens to help anything extracting a summary. It costs nothing and it is good writing.
Publish things that cannot be summarised from elsewhere
Your own data. A test you ran. Prices you verified. Anything derived from existing content can be assembled without you.
Be consistent about the facts of your own business
Your name, what you do, your pricing model, stated the same way across your site and anywhere else you appear. Inconsistency is a mundane reason to be described wrongly.
Make your pages readable by machines
Clean headings, real text rather than text inside images, working structured data where it applies. This is ordinary technical hygiene with a long history.
Get mentioned in places that get read
Being named in comparisons, roundups and industry writing has always mattered. It plausibly matters more when models summarise the discourse rather than the page.
Notice what is missing: any tactic that only makes sense if a specific theory about model behaviour is true. That is deliberate.
Nobody controls what a model outputs, and outputs vary between runs for the same prompt. A guarantee here is either a misunderstanding or a lie.
Publishing volume as the strategy
Pumping out thin pages to increase the chance of being quoted is the scaled-content pattern that gets whole domains suppressed in ordinary search. You would be trading a known asset for an unproven one.
Tools charging to track your mentions
Some are genuinely useful. Ask how they sample, how often, and how they handle the fact that the same question returns different answers on different runs. If they cannot explain their sampling, the dashboard is decorative.
Anyone certain about ranking factors
The systems are unpublished and change frequently. Confidence here is a sales posture, not a finding.
Segment referrals from assistant domains in your analytics. Imperfect, because plenty of assistant-influenced visits arrive with no referrer at all, but it gives you a floor rather than a guess.
Watch for the ranking-holds-clicks-fall pattern. A page keeping its position while losing clicks is the clearest signature of an answer being given without you.
Ask new customers how they found you. Unfashionable, unscalable, and currently more reliable than any tracking for this specific question.
Track branded search volume. If being cited is doing anything for you, more people looking you up by name is where it shows.
The framing that keeps you sane
Treat AI visibility as a distribution channel that is early, badly measured, and not yet worth reorganising around. Do the cheap things that also serve readers. Refuse to buy anything that only pays off if an unverified theory holds. Revisit in six months, by which point somebody may have published a study with a methodology you can actually inspect.
Figures on conversion lift from AI referrals are cited here as examples of the spread in published claims, not as evidence of any particular number. They come from vendor and agency analyses of their own or their clients' data, with methodologies that are not directly comparable. Two illustrative examples:
We were unable to trace any of these to a study with an inspectable methodology and an independent measurer, which is the reason they are presented as contested rather than as findings.
