A lot of teams talk about AI visibility like it starts on somebody else's website. More Reddit. More press. More mentions. More everything.
That is usually backwards. The answer should be easy to quote before you chase the mention.
PostHog's new AEO write-up is useful because it treats the problem less like a magic trick and more like content, attribution, and judgment work.
Start with a reality check, not a vendor deck
PostHog private-mode three-prompt benchmark before AEO tool sprawl is the clean opening move. Ask the models the boring buyer questions first. See whether the product shows up, disappears, or gets described badly.
That sits well beside Fern API catalog before agent scrape guesswork. Before the team buys more reporting, it should make the product legible enough to be found at all.
Pages need to survive extraction
PostHog citable content chunks before monolithic SEO pages is the core lesson. The page still matters, but the quoteable unit is smaller now. A heading, a definition, a table row, a specific number.
I would read that with llms-full single-file context export and well-known llms aliases for agent compatibility. The easier it is to retrieve the right chunk, the less guesswork the model has to do.
Clean your own house before chasing every side street
PostHog owned-surface cleanup before off-domain AEO sidequests is the part most teams want to skip because it sounds ordinary. Ordinary is the point. Fast pages, direct language, healthy internal links, and obvious routes still do most of the work.
This belongs next to custom docs 404 page with task-led redirects and recurring docs audit for broken links and style drift. A citation is wasted if the next click lands on fog or friction.
Ask the buyer what they actually typed
PostHog onboarding prompt capture for AI attribution is my favorite move in the batch. The growth team does not need another synthetic prompt brainstorm if real converting users are willing to hand over the sentence that brought them in.
That is high-signal language. It is better than a Slack debate. Better than a vendor screenshot. Better than a brainstorm where everybody pretends to be the customer for twenty minutes.
Reporting is stitched, not discovered whole
PostHog AEO reporting quilt before single-source certainty matters because AEO is still messy. One source tells you what was crawled. Another tells you what got said. Another tells you what converted. None is enough alone.
That fits with Fern search scope by product before cross-doc noise. Retrieval systems get worse when the team mistakes one partial view for the whole map.
Be willing to throw away a flattering prompt set
PostHog synthetic prompt reset when real user language disagrees is the discipline most teams avoid. If the tracked prompts do not resemble what real users say, the history is comforting but useless.
This is strongest for SaaS, AI products, developer tools, marketplaces, and creator software where the buyer often arrives with a question before they arrive with brand loyalty.
If I were tightening one AEO program this week, I would benchmark a few plain prompts, rewrite the owned pages into quoteable chunks, ask AI-sourced signups what they actually typed, stitch together a report from multiple imperfect sources, and scrap any prompt set that no longer sounds like a real buyer. That is slower than hype and much more useful.
If you want help turning docs, answer surfaces, and AI discovery into a cleaner acquisition path, the advisory CTA is here: work with Ian Goh.