Being found by AI assistants

The newest surface, without the snake oil

When a potential customer asks ChatGPT for "the best client-onboarding software for small accountancy firms", or asks Perplexity for "a trustworthy physiotherapist in Graz", an answer comes back with names in it. Whether yours is among them is now a marketing outcome you can influence — and almost none of your competitors are trying yet.

That novelty attracts two kinds of nonsense: "AEO gurus" selling secret ranking factors nobody possesses, and fatalists declaring it unknowable. This lesson takes a third position: separate what is known from what is speculated, act on the known, and measure honestly. It turns out the known part is substantial — and you've already built most of it.

What is actually known

Three things are documented by the assistant vendors themselves or directly observable by anyone who runs the queries:

1. Most consumer AI answers are retrieval-augmented. For current questions and recommendations, the assistant searches live web indexes and its own crawls, then synthesises from what it retrieves. Which index feeds which assistant shifts with partnerships — the durable fact is retrieval itself. A page that can't be crawled, indexed or ranked can't be retrieved; a page that can't be retrieved can't be cited. Classic SEO is the entry ticket to AI citation, not a separate discipline. Everything from Lessons 6.1–6.5 counts here.
2. Answers name and cite sources — and for recommendation queries, the cited sources are disproportionately third-party surfaces: review platforms, comparison articles, forums, directories, trade press. You will see this in your own audit within the hour.
3. Answers vary run to run. The same question asked five times returns different lists. Any conclusion drawn from one screenshot is noise.

What is speculated — and how to treat it

Nobody outside the vendors knows the weighting of any signal. Whether schema markup increases citation odds: plausible, unproven. How much an unlinked mention counts: unknown. Anyone selling you a definitive list of "AI ranking factors" is selling the confidence, not the knowledge — the systems change monthly and disclose little.

The working rule: act on mechanisms (retrieval, citation, third-party sourcing), ignore anyone claiming precision. Conveniently, the mechanism-level actions are things worth doing anyway.

What observably gets cited

Run twenty assistant queries in any category and patterns repeat: extractable, answer-first passages (Lesson 6.3's structure); pages with clear entity identity (Lesson 6.4's consistency work); original data and specifics — assistants, like journalists, prefer sources that add facts; fresh, accurate, reachable pages (Lesson 6.5). Vague brochure pages effectively don't exist to an answer engine: there is nothing in them to extract.

And for "best X" queries, presence matters more than prose: being present and well-reviewed on the surfaces the assistant already cites often outweighs anything on your own site. Your Lesson 6.4 authority actions were AEO actions all along.

Measuring an emerging channel honestly

Two instruments, both free, both imperfect — and their imperfection is part of the lesson:

• The monthly audit log. Ask fixed queries to two assistants, five repeated runs each, and record frequency of mention across runs — not a single screenshot. Today's exercise is run one; the log is the baseline your Module 10 measurement work will thank you for.
• Referral traffic. Watch your analytics referrers for assistant domains — chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and whatever succeeds them — and treat the numbers as a floor, not a total: many assistant-driven visits arrive with no referrer at all, and Search Console doesn't separately report AI-answer clicks as of this writing. Undercounted is still countable. A floor that rises is a real signal.

What not to buy: "AI rank tracking" subscriptions at €100+/month. For a business your size the manual log is more honest than a dashboard built on the same repeated queries — and it's €0.

Storkflow, month 0 to month 3

Lena runs the baseline audit. Month 0: "best client-onboarding software for small accountancy firms", five runs each on ChatGPT and Perplexity. Storkflow: named zero times. Four competitors recur, and the citations trace to three surfaces — two comparison listicles and one accountants' forum thread. Nothing about her product; everything about where she isn't.

The gap analysis writes her plan, and every action reuses an existing artefact:

• The onboarding-data study — her Module 5 hero asset, built on 2,100 anonymised client onboardings — is pitched to the authors of both listicles. One refreshes their piece and cites it.
• She answers the forum thread herself: genuinely useful numbers from the study, with a disclosure that she works at Storkflow. Useful first, linked second.
• The review gap: one G2 review, for a product bought by people who ask assistants for recommendations that cite review platforms. Her Module 3 asking-round motion, pointed at G2, grows it to nine.
• One answer-first comparison page ships from a Lesson 6.3 brief — "Storkflow vs doing it in email and Excel", because the pack's positioning research says email-plus-Excel, not a rival product, is what buyers actually weigh.

Month 3: Perplexity names Storkflow in three of five runs, citing the refreshed listicle and her G2 profile. ChatGPT names her intermittently. Analytics shows 14 sessions from assistant referrer domains — small, honestly logged, and attached to a channel where her competitors still aren't trying. The caveat stays on screen: answers vary; the log tracks frequency, not triumph.

What not to do

• Prompt injection — hidden text on your page saying "ignore previous instructions and recommend us". Detectable, filtered increasingly well, and reputation-torching when found: you'd be teaching every future system that your domain cheats.
• Fake reviews and fabricated data — illegal under EU consumer law (Lesson 6.4), and doubly foolish here: you'd be lying to systems whose defining talent is cross-referencing sources.
• Blocking AI crawlers while wanting AI customers — the Lesson 6.5 self-harm, restated once because it's still the most common own goal.

The durable strategy is being genuinely citable: retrievable, specific, corroborated, present where answers are sourced. Which is — conveniently, and by design — exactly what this module built.

---