Does Your SEO Agency Actually Understand AI Search? Five Questions You Can Verify Yourself
Since AI search became a topic, AEO and GEO have appeared on nearly every provider's site. The trouble is that those two terms are now where “we do SEO” was a decade ago — saying it costs nothing and proves nothing. These five questions aren't designed to catch anyone out. They're here because each has a reasonably settled answer, and the wrong answers tend to be wrong in obvious ways. You don't need to be technical to tell the difference.
One: does llms.txt help with Google's AI features?
This is the fastest way to tell whether someone is following the discourse or reading the primary sources.
The answer is no. Google's own AI-optimization guidance states that you don't need to create machine-readable files of this kind to appear in Google Search, including its generative features, and says plainly that Google Search itself doesn't use them.
If a provider tells you that adding an llms.txt will lift your AI visibility, that isn't a technical judgement — it's someone else's marketing, repeated. The file is harmless, and it has genuine uses as developer documentation, but it is not a visibility lever with Google.
Note what kind of question this is: the answer sits in Google's published documentation and you can check it yourself. A provider actually reading primary sources will tell you it doesn't work, even though admitting that does them no favours.
Two: how do you measure AI citations?
The honest answer is that you can't measure the whole of it.
Search Console reports impressions and clicks in search results, but there is no equivalent for ChatGPT, Perplexity or Gemini telling you how often people asked about your brand this month or how often an AI cited you. Tools exist that sample prompts and record what comes back — that is sampling, not complete data.
So the reasonable approach is to measure what can be measured — organic performance, whether content is being indexed correctly, brand mentions — and to say plainly which parts can only be observed rather than quantified.
If a provider produces a dashboard claiming full tracking of AI citations, it's worth asking where the data comes from, how many samples, and which platforms. The problem isn't that such tools are worthless; it's presenting a sample as a complete picture.
Three: can AI crawlers reach my site right now?
This question separates whoever handles the technical base from whoever only writes content.
For an AI tool to cite you, it has to be able to reach you. And the reason it can't is often not robots.txt but protection settings: CDN or WAF rules drawn too broadly will treat search engines and AI crawlers as anomalous traffic and block them along with everything else.
The awkward part is that this failure has no symptoms. The site loads normally for visitors, speed is fine, nothing errors — but crawlers receive a block. We've handled a real case of it: a clinic site applied blanket blocking after a DDoS attack, which stopped the attack and shut Googlebot out with it, and the site drained out of search results.
A sound answer will mention verifying as a crawler rather than checking in a browser. If it has never occurred to a provider that protection rules affect crawling, that layer probably isn't part of what they do.
Four: does AEO do anything if the content isn't indexed?
It doesn't. The question is worth asking because it reveals the order in which someone understands the work.
Google's AI Overviews and AI Mode are grounded in the same ranking systems as classic Search. A page has to be crawlable, indexed, and eligible for snippet display before it can appear in any AI feature at all. Indexation is the floor, not a bonus.
So if a site's content isn't being indexed properly, adding question formatting and structured data changes nothing — it optimises the shape of a page nothing is reading. The order is crawlability and indexation first, content structure second.
It's the same reason we run the architecture scan before touching content. The sequence isn't a preference; it's causation.
Five: will you guarantee that AI cites us?
Any provider who answers “guaranteed” can be ruled out on the spot.
Whether an engine cites you is decided by its model. Models update, source-selection changes, and no outside provider controls it. What can be done is the structure, the answer formatting, entity consistency and authority signals — work toward being citable. That's a direction of effort, not a promise.
The same standard applies to rankings. A company willing to tell you what it doesn't guarantee while it's still selling to you will usually report honestly once you're working together. One that will hand you a guarantee before a contract tends to explain the data the same way afterwards.
The value of this question isn't the answer. It's watching what someone does with an opportunity to overstate.
Three things you can check yourself
No tools required, and no need to take anyone's word for anything.
First, open the provider's own site and see whether it does what they're selling you — structured data, clear question-and-answer blocks, performance claims you can check. A company that specialises in search visibility has published its own working sample.
Second, look at whether their published figures state a measurement period, a baseline, and a source. “Traffic up 300%” with no starting point, no window and no source is a sentence, not evidence.
Third, ask whether they've had engagements that didn't work, and what happened. A provider willing to describe limits and failures is usually more accurate about successes too.