Essay — — by Mahmoud Zalt
A reflection on what actually makes a tool recommendable inside ChatGPT and AI Overviews: passage citability, brand mentions on Reddit and YouTube and Wikipedia, and clear side-by-side comparison content.
One afternoon we noticed a new signup with a referrer string we had not seen before. ChatGPT. The person had asked something close to "what is a good AI workforce platform for a solo founder", and our name had come up in the answer. They clicked the link, read one page, and signed up.
That was the first time we took AI search seriously as a real channel. We started asking the same questions ourselves and reading what the model was actually citing. Some of it was us. Most of it was not. The pages that did get pulled in had a pattern, and it was not what we expected.
Google ranks pages. AI search pulls passages. That sounds like a small distinction and it is not. ChatGPT and AI Overviews stitch together short blocks of text that already read like a clean answer. A page can rank well on Google and still be useless to an AI engine if every paragraph buries the answer under two sentences of setup. The model reaches for whatever can be quoted as-is.
There is a second layer. The model does not trust a passage just because it exists. It cross-references the brand against forum threads, video transcripts, encyclopedic entries, and comparison roundups. A tool with mentions on Reddit, a few YouTube reviews, and a stub on Wikipedia or a structured directory shows up as an entity the model can confidently name. A tool that lives only on its own website tends to get cut in favor of something more verifiable. The recommendation is not about quality alone. It is about being legibly real outside your own marketing.
Every section now opens with the actual answer in the first sentence, then explains. We rewrote H2s to match how people phrase questions in ChatGPT ("What is the cheapest AI workforce platform", "Which AI employee tool works without code"). The opening sentence is the citation target. This single change moved more pages into AI Overview citations than any keyword work we had done before.
Side-by-side tables with named competitors and concrete values (setup time, hosting model, included credits, supported channels) get cited far more than prose. AI engines lift table rows almost verbatim. We added one comparison table per alternative page and saw those pages pulled into answers within weeks. Tables also force you to be honest, because abstract claims look thin next to real values.
We stopped treating off-site mentions as a vanity metric. A few honest Reddit answers, a couple of YouTube walkthroughs, and listings on a few directories shifted how often the model named us by name. Without those outside signals, the model describes the category and skips the brand. With them, the brand becomes a recognized entity instead of an unknown string.
Our old paragraphs ran six or seven sentences and buried the answer in the middle. We cut them down so each one stands alone as a quotable block. If a reader (human or model) only reads that paragraph, they still walk away with a clear answer. The most boring change on the list, and the one with the largest effect on pickup rate inside AI Overviews.
Before any of the work above, we chased a handful of tactics that sounded clever and went nowhere. We added an llms.txt file because blog posts insisted it was the new sitemap for AI. The major engines have not committed to reading it, and our logs show no crawlers honoring it. We tried writing in an "AI-friendly" tone with awkward keyword stuffing aimed at vector similarity, which mostly made the prose worse. We obsessed over content chunking before realizing the models do their own chunking and what they actually want is paragraphs that already read as answers. Google itself has been clear that for AI Overviews, classic search quality signals still apply: helpful content, clear structure, a recognizable entity. The novelty advice was loud. The unglamorous fundamentals moved the needle.
AI search does not reward the cleverest content. It rewards the most quotable paragraph attached to a brand the rest of the web has already named.
It is easy to read the GEO discourse and conclude that classic SEO is over. It is not. A large share of citations in AI Overviews (Google's own analysis put it around ninety-two percent) come from pages already ranking in the top ten organic results for the underlying query. If you are not ranking, you are not being cited. The AI layer sits on top of search, it does not replace it. The boring work still matters: a crawlable site, internal linking, real backlinks, clear page titles, and content that earns its position. What changes is the writing style on top of that foundation. Skipping the base layer and chasing only "GEO tactics" is how teams end up with thin pages the model still does not cite.
We are slowly converting older long-form posts into the same shape: question-led H2s, first-sentence answers, one honest comparison table per page, shorter paragraphs throughout. We are also more deliberate about the off-site footprint. Not paid placements, just being genuinely present where founders compare tools (a handful of subreddits, a few directories, a Wikipedia stub when independent coverage justifies one). AI engines change their citation logic constantly, so we do not treat this as a one-shot fix. The bet is simpler: write so a single paragraph answers one real question, and be visible enough across the open web that the model trusts the name.
Partially. The crawl layer, ranking signals, and backlink fundamentals are the same: if you are not ranking on Google, you are unlikely to be cited in AI Overviews. What differs is the writing layer. AI engines pull short, self-contained passages, so paragraph structure (answer first, context after) matters more than it ever did. Think of GEO as SEO plus a stricter editorial style, not a replacement.
Probably not as a priority. The major AI engines have not publicly committed to reading llms.txt, and our logs show no consistent crawler traffic respecting it. Adding one is harmless housekeeping, not a growth lever. The time is better spent rewriting important pages so each paragraph stands on its own as a clear answer.
Faster than classic SEO, slower than ads. The first AI-search referrals showed up within a few weeks of rewriting our most-trafficked pages around question-form H2s and first-sentence answers. Comparison tables get cited even faster because the model can lift rows directly. Off-site work (Reddit, directories, video reviews) compounds slowly over months. No instant lever, only an accumulating one.
The thread running through those questions, and through everything we changed, is simpler than it first appears. Being recommended by a model is not a trick of formatting or a clever tag in the head of a page. It is the cumulative result of being legible: writing that can be read in a single passage, a brand that is named in places the model already trusts, a footprint that survives even when nobody is searching for you by name. The same idea shows up in a quieter form in another essay we wrote about being visible to a model in the same way you would want to be visible to a teammate or a user. The two topics rhyme more than we expected when we started writing them.
Looking back at the months we spent on this, the lesson that keeps surfacing has nothing to do with AI search as a tactic. It is that the work which moves the needle is almost always the work that felt too obvious to do. Rewrite a paragraph so it answers a question on its own. Put a real comparison table on a page that needed one. Show up in three or four places outside your own site, with your real voice, because that is where people (and models) check whether you exist. None of it is novel. None of it is the kind of advice that goes viral on a Tuesday. And yet every time we tried to skip past the fundamentals in pursuit of something more interesting, we lost weeks. What this all really comes down to is patience with the slow, unfashionable layer of the work, and a quiet trust that being legibly useful, in public, is still the cheapest distribution any of us will ever have.