On this page 11 sections
Assistants name companies from two places: what was in the model’s training data, and what it retrieves at answer time. You cannot edit the first. You can influence the second - by existing in the third-party sources assistants pull from, and by being the clearest answer on pages they can actually read. Most invisibility is a sourcing problem, not a ranking one.
Before spending anything on this, it is worth being clear about how much of the mechanism is genuinely understood and how much is inference. A lot of what is sold as expertise here is the second dressed as the first.
Key takeaways
- Two mechanisms, only one of which you can influence in a useful timeframe.
- Assistants lean on third-party lists, comparisons and directories far more than on your own site.
- Being cited and being recommended are different outcomes with different causes.
- Answers vary between runs, so a single test tells you nothing.
- The honest measurement is a sampled panel of prompts over time, not a rank.
Who this applies to
You have asked an assistant a question your business should be an answer to - “best AI automation agency for mid-market ecommerce” - and it named three competitors and not you.
The two mechanisms
What the model already learned
Training data is a snapshot. If your company was widely written about before the cutoff, the model has some representation of you; if not, it does not.
You cannot change this retroactively, and no vendor can. Claims of getting you “into the model” are not credible. What you can do is build the kind of third-party presence that lands in future training data, which is a slow, real strategy with a payoff measured in years.
What it retrieves at answer time
The part you can actually work on. When an assistant searches the web mid-answer, it runs queries, reads results, and synthesises. That pipeline has a shape you can be present in - and it usually leans on the same pages that rank in conventional search.
This is why the honest version of most advice here overlaps heavily with ordinary SEO. That is not a reason to dismiss it. It is a reason to be suspicious of anyone selling it as an entirely new discipline.
Why it names competitors
Four reasons, roughly in order of how often they are the real one.
They are in the lists. Ask about vendors in a category and an assistant frequently reaches for roundups, directories and comparison pages - “top 10 X agencies”, G2, Clutch, industry association listings. If you are not in those, you are not in the retrieved set, regardless of how good your site is.
This is the most common cause and the most fixable. It is also unglamorous: it is directory submissions and being included in other people’s comparisons.
Their pages answer the question directly. A page that states an answer in its first paragraph is easier to extract than one that builds to a conclusion over eight paragraphs of context. Structure affects extractability.
They have more third-party corroboration. Reviews, mentions, case studies on other domains, conference listings, podcast appearances. Assistants weight corroborated claims, and a claim that exists only on your own site is uncorroborated by construction.
You are not clearly in the category. If your site describes what you do in language nobody searches for, you may be invisible for a vocabulary reason rather than an authority one. Check whether your service pages use the words a buyer would use.
Cited, mentioned, recommended
Three different outcomes, frequently conflated by tools that sell “AI visibility”:
- Cited - your URL appears as a source. Driven by having a page that directly answers the query and is retrievable.
- Mentioned - your name appears in the text. Can happen without a citation, from training data or from another page that names you.
- Recommended - you are put forward as a good choice. Needs corroborating signals: reviews, comparisons, third parties saying you are good.
They have different causes and different fixes. A vendor reporting a single “visibility score” is flattening three things that need separate work.
What you can actually influence
In descending order of how confident anyone should be about the effect.
1. Third-party presence. Directories, review platforms, industry lists, other people’s comparison articles, podcasts. High confidence - this is where the retrieved sources come from, and it is verifiable by looking at what gets cited today.
2. Answer-first page structure. Lead with the answer, then support it. Clear headings that match real questions. Self-contained sections that survive being extracted from context. Reasonable confidence, and it costs nothing because it also improves the page for humans.
3. Being crawlable by AI crawlers. Assistants use their own user agents. If your robots.txt blocks them, or your content requires JavaScript to render, you are absent for a mechanical reason. Verifiable, and worth checking first because it is binary.
4. Structured data and machine-readable summaries. Lower confidence. Schema.org markup demonstrably helps conventional search; whether it materially changes assistant behaviour is not well established publicly. It is cheap and low-risk, which is a different argument from it being proven. See llms.txt and structured data.
5. Content volume. Lowest. Publishing more pages does not make an assistant recommend you if none of them are the clearest answer to anything, and thin content is arguably worse than none.
How to actually check
Do not test once. Answers vary between runs for the same prompt, and a single query is anecdote.
A workable baseline, roughly an hour to set up:
- Write 20-30 prompts a real buyer would use. Category questions, comparison questions, problem-first questions. Not your brand name - that tells you nothing about discovery.
- Run each three times, across the assistants your buyers use.
- Record three things per run: were you mentioned, were you cited, were you recommended.
- Record who else appeared, and where those answers were sourced from. The source list is the most actionable output - it tells you exactly which pages to be present in.
- Repeat monthly. You are watching a rate over time, not a position.
That is a sampled panel, and it is genuinely how this gets measured. Anyone offering a precise rank for a query is over-claiming: there is no equivalent of Search Console here, and the underlying system is non-deterministic.
What we tell clients about this category
Two things, and the first costs us work.
A large share of what is sold as GEO is repackaged SEO with a new invoice. Be answer-first, be crawlable, get cited by third parties, use clear structure - all of that was good advice before assistants existed. Where a vendor cannot explain what is genuinely different about their approach, the honest assumption is that nothing is.
The measurement is weaker than the marketing. We report a sampled share of citations across a prompt panel, with the run-to-run variance stated, because that is what the mechanism supports. It is a less impressive artefact than a dashboard with a single number trending upward, and it has the advantage of being true.
Where we do think there is something real: the sourcing layer. Which third-party pages an assistant reaches for in your category is observable, it differs by category, and being absent from those pages is a concrete, fixable gap that most businesses have never looked at. That part is worth paying for. The rest is good content practice with a new name.
We also tell clients the uncomfortable version: if your domain has no authority and no third-party presence, this is not the first thing to fix. It is downstream of having something worth citing.
When this is not worth working on
When nobody discovers you this way. If your pipeline is referral and outbound, assistant visibility is a vanity project. Check before investing.
When you have no third-party presence at all. Start with review profiles and directories - that is the input to the mechanism, not an alternative to it.
When the category is too new to have consensus sources. If assistants have nothing good to retrieve for your category, they improvise, and nothing you do to your own site changes that quickly.
When you would be optimising for a metric nobody has agreed. Decide what outcome matters - cited, mentioned, recommended - before buying a tool that measures a fourth thing.
Frequently asked questions
Can we pay to appear in AI answers?
Not in the organic answer itself, at time of writing. Advertising products in and around assistants are appearing and are separate from the organic mechanism. Anyone offering guaranteed organic placement is describing something that does not exist.
How long does any of this take?
The crawlability and structure work shows up in weeks if it shows up. Third-party presence takes months. Training-data representation takes years, if at all.
Do we need to allow AI crawlers?
It is a trade. Blocking them protects content from being used without attribution and guarantees absence from retrieved answers. Most businesses selling services should allow them; publishers monetising pageviews reasonably decide otherwise.
Is there a Search Console equivalent?
No. Some providers surface limited referral data, and third-party tools sample prompts the way described above. Treat any precise ranking claim with suspicion.
Does our own site matter at all?
Yes, for being cited once retrieved, and for the vocabulary question. It matters less than most site owners assume relative to third-party sources.
Next step
The most useful first step is finding out what is actually being cited in your category, which takes about an hour and requires no tooling. The AI search visibility engagement does that as a baseline, with the variance stated rather than hidden.
Related: Writing content that assistants actually cite · llms.txt and structured data · GEO or SEO: what actually changes · AI search visibility