Every SEO newsletter in 2026 has a GEO section. Most of it is recycled 2015 content-marketing advice with "AI Overviews" pasted on top, plus a new file called llms.txt that almost nothing actually reads yet. Some of it is genuinely useful. Telling the difference starts with understanding how these systems actually retrieve content — not how a LinkedIn post says they do.
Answer engines still need an index
AI Overviews sits on top of Google's existing index. ChatGPT's web search has relied on Bing's index plus OpenAI's own retrieval crawler, OAI-SearchBot — a separate crawler from GPTBot, which OpenAI uses for training data and which blocking does not remove you from ChatGPT's search results. Perplexity runs its own crawler, PerplexityBot, for indexing, plus a separate user-triggered agent for live fetches. None of these systems answer questions purely from what got baked into model weights during training — training cutoffs run months to a year behind, and nobody wants an assistant that can't discuss this week's news. For anything specific or time-sensitive, the model retrieves live pages, extracts passages, and synthesizes an answer with citations.
That means the first requirement for GEO is identical to the first requirement for ordinary SEO: your page has to be crawled, rendered, and indexed correctly. If it isn't, no amount of "AI-optimized" content structuring matters, because nothing ever reaches the model in the first place.
The crawler compatibility gap that actually matters
Here's the part most GEO advice skips: OAI-SearchBot, PerplexityBot, and ClaudeBot do not reliably execute JavaScript. Googlebot has run a two-wave rendering process — fetch the HTML, queue it for headless rendering, index again — for close to a decade. The AI crawlers, as of 2026, mostly fetch raw HTML and stop there.
If your pricing, specs, or comparison content loads via a client-side fetch call after hydration — common on single-page apps and plenty of framework defaults — an AI crawler sees an empty shell and a loading state, not your content. The page can rank fine in Google, because Googlebot waited around for the render, and still be functionally invisible to ChatGPT search or Perplexity. The test is simple: view-source on a key page, and if the numbers and claims your sales team cares about aren't sitting in the raw response, an AI crawler that doesn't render JavaScript never saw them either.
The fix isn't exotic. Server-side rendering, static generation, or a pre-rendered snapshot served to bot user agents gets you back on even footing. This is engineering work, not a content trick — which is exactly why it gets skipped by advice written from a marketing seat rather than an engineering one.
Passage-level clarity beats keyword density
Classic SEO optimized for a whole page ranking on a whole query. Answer engines extract a passage — commonly a short paragraph, not a full page — and cite or paraphrase it directly. That changes what "good content structure" means:
- Put the direct answer in the first sentence or two under a header, not at the end of a three-paragraph windup.
- Make each section self-contained. A paragraph that only makes sense after "as discussed earlier" is a poor extraction candidate.
- One clear claim per sentence when you're stating a fact, number, or definition. Dense, hedge-everything B2B prose is harder for a model to lift cleanly than a plain declarative sentence.
- Lists, tables, and defined terms extract more reliably than the same information buried in narrative paragraphs.
None of this requires keyword stuffing. Most production retrieval — Google's included — is hybrid, combining keyword-level matching with embedding-based semantic search and neural re-ranking, so repeating a phrase five times to "signal relevance" does nothing except make the passage worse to extract without actually improving the match.
Where structured data still earns its keep
Schema markup — Article, Organization, Product — doesn't guarantee a citation, but it does disambiguate what a page is about for Google's index, which is what AI Overviews draws from; there's no public confirmation that OpenAI's or Perplexity's own retrieval systems weight JSON-LD the same way. It costs almost nothing to implement correctly regardless. (FAQPage is worth a separate note: Google removed FAQ rich results from Search entirely in 2026, so that specific type no longer earns a visible result anywhere — it can still help entity clarity, but not a rich snippet.) Treat it as infrastructure, not magic. The same goes for basic E-E-A-T signals: a visible author, a publish or last-updated date, links to primary sources. Answer engines weigh these as trust signals because citing an unattributed, undated page carries hallucination risk for them — and reducing that risk is their product problem as much as yours.
llms.txt: a hedge, not a strategy
llms.txt — a proposed markdown file at your domain root summarizing key pages for LLMs to consume — gets treated in a lot of GEO content like a new robots.txt that unlocks AI visibility. As of 2026, none of the major AI providers have published documentation confirming their crawlers or retrieval systems parse and prioritize it as a ranking or citation signal. It costs an afternoon to add, and it won't hurt anything. But when a vendor sells "llms.txt optimization" as a headline deliverable, that's the cargo cult in action: a low-effort file marketed as a strategy because it's easy to sell and hard to disprove.
Myths worth retiring
A few more patterns worth dropping this year:
- "Optimize for the model's training data." New content doesn't enter a shipped model's weights. Citation happens through live retrieval, not training inclusion.
- "AI Overviews are killing organic, so stop doing SEO." Zero-click behavior is real and worth planning around, but the answer is fixing crawlability and structure — not abandoning fundamentals, which just makes you invisible everywhere at once.
- "There's a proprietary AI-visibility score you should chase." Most of these tools sample a handful of prompts against a black-box methodology. Treat any single "GEO score" as directional at best, not a KPI.
What to actually do this quarter
- Confirm your key pages serve full content in raw HTML, not only in the rendered DOM.
- Check robots.txt for GPTBot, PerplexityBot, ClaudeBot, and Google-Extended, and set access on purpose rather than by default.
- Rewrite your top 10-20 pages so each section answers its own header in the first sentence.
- Add Article, FAQ, or Organization schema where it's genuinely accurate — not padded in for the sake of it.
- Add llms.txt if you have a spare afternoon. Don't build a quarter's roadmap around it.
- Start tracking referral traffic from chatgpt.com and perplexity.ai in analytics now, even at low volume — it's the leading indicator for whether any of this is working.
Bottom line
GEO in 2026 isn't a new discipline bolted onto old SEO. It's mostly the technical-SEO fundamentals that were always worth doing — server-rendered content, clean information architecture, real structured data — applied to a slightly different set of crawlers and consumption patterns. The content-strategy layer changes less than the vendors selling "AI search optimization" packages want you to believe.
If you're not sure whether your site is even reaching these crawlers correctly, that's the question to answer before spending another dollar on content. Our SEO team runs crawlability and passage-structure checks as part of every technical audit — get in touch if you want a straight answer on where your site actually stands.
