How to Rank in ChatGPT Search Results

• ChatGPT mentions come from two different mechanisms — training memory and live search — and they respond to different levers, on different timescales.
• If GPTBot or OAI-SearchBot is blocked, none of the rest of this matters — you've opted out before the game starts.
• Being quotable to humans (clear claims, named numbers, definitions) is the same asset that makes you quotable to a model.
• Third-party mentions compound — one listicle naming you does little, twenty independent ones over a year change what the model "knows".
• Narrow, specific prompt spaces are winnable now; broad category prompts are a multi-quarter authority game.
• None of the shortcuts — stuffed phrases, hidden prompts, mass AI content — survive contact with how these systems actually filter sources.

How ChatGPT decides who to mention

Two paths lead into an answer, and conflating them is the most common strategy mistake I see. Training-data memory is what the base model learned during pretraining and subsequent fine-tuning — it’s static between model versions and it’s why some brands get named even for prompts with no live search triggered at all. Live search retrieval is the search layer (built on Bing indexing plus OpenAI’s own crawl via OAI-SearchBot) fetching and ranking pages at query time, then citing what it used. Ask ChatGPT something time-sensitive — pricing, “best X in 2026”, recent news — and you’re almost always in retrieval territory. Ask something conceptual or definitional — “what is GEO”, “who does X kind of work” — and you’re often pulling on memory, shaped by how consistently the web already talks about you.

The practical test I use with clients: open a clean session, ask the target prompt, and check whether the response reads like it’s citing specific pages (retrieval) or speaking in generalities without visible sourcing (memory). That single check tells you which lever to pull — chase crawlability and on-page structure for retrieval problems; chase third-party mentions and consistency for memory problems. Treating a memory problem as a technical SEO problem wastes a quarter.

The quick wins (weeks)

• Let it in: allow GPTBot and OAI-SearchBot in robots.txt, and check server logs, not just the robots file — a CDN, WAF or bot-management rule can silently 403 these agents even when robots.txt says yes. I’ve seen this exact mismatch cause a client to be functionally invisible for months while their robots.txt looked perfectly permissive.
• Answer-shaped pages: definitional blocks, question headings, direct first sentences — the same AEO mechanics. Put the answer in the first sentence after the heading, not the third paragraph. Retrieval systems chunk pages; if your answer sits below the fold of the chunk, it may never get pulled into the citation.
• Server-rendered content: if your value only exists after JavaScript runs, assume the crawler never saw it. Test this directly — fetch the URL with a plain HTTP request, no JS execution, and read what comes back. If the key claim isn’t in that raw HTML, it isn’t reliably in the model’s retrieval either.
• One clean canonical claim per page: pages that try to say everything get quoted for nothing. A page that states one number, one definition, one recommendation clearly is far easier to lift into an answer than a page that hedges across five subtopics.

The slow wins (months)

• Entity consistency: one name, one description, everywhere — site, LinkedIn, directories, podcasts, press mentions. Inconsistent descriptions (“SEO agency” here, “digital marketing consultancy” there, “growth studio” elsewhere) split the signal the model uses to resolve who you actually are. Pick the description, use it verbatim for a year, then reassess.
• Third-party mentions: reviews, listicles, comparisons naming you. Models trust what many independent sources repeat — a single glowing self-published case study carries far less weight than five unrelated third parties describing you the same way. This is why digital PR and genuine industry mentions matter more for GEO than for classic backlink SEO — you’re not chasing link equity, you’re chasing corroboration.
• A citable claim: numbers, definitions or frameworks other sites quote. Being quotable by humans is the training set for being quotable by models. If nobody has ever quoted your definition of a term, ask why — usually it’s because the definition is vague, or buried, or indistinguishable from ten competitors’ versions.
• Consistency over time, not a burst: a spike of ten mentions in one month reads differently to an algorithm trained on years of web data than the same ten mentions spread across two years. Slow, steady presence beats a single PR push, however large.

Training memory vs live retrieval: how to tell which one you’re fighting

This distinction deserves its own section because most “how to rank in ChatGPT” advice quietly assumes one and applies it to both. If your prompt space is stable and evergreen — “what does a GEO consultant do”, “define answer engine optimization” — you’re fighting for a place in training memory, and the fix is months of consistent, corroborated presence across the web, not a page edit. If your prompt space is comparative or current — “best GEO agencies 2026”, “who tracks ChatGPT visibility” — you’re fighting for retrieval, and the fix is crawlability, structure and freshness on your own pages. Most brands need both, but spending retrieval-fix budget on a memory problem (or vice versa) is the single most common wasted quarter I see in this space.

How long does it actually take to show up in ChatGPT answers

For retrieval-driven prompts, weeks — once GPTBot and OAI-SearchBot can reach and parse the page cleanly, citation can happen on the next crawl-and-index cycle. For memory-driven prompts, there’s no shortcut timeline: it depends on the training and fine-tuning cadence of the model version in use, and on how much of the wider web has already reinforced the association before you started. I set client expectations in two tracks — a “can we get cited this quarter” track for narrow, specific, retrieval-friendly prompts, and a “are we building the association” track for broad category prompts, measured over two to four quarters, not weeks. Selling the second on the timeline of the first is how GEO programmes lose internal credibility in month two.

Do llms.txt and structured data actually help

Structured data (schema.org markup — Organization, Article, FAQPage) has a defensible case: it makes entity facts machine-parseable and reduces ambiguity, which helps both classic search and AI retrieval layers resolve who you are and what a page claims. I implement it as standard practice. llms.txt — a proposed convention for pointing language models to a curated summary of a site — is different: no major AI vendor has committed to reading it as a ranking or retrieval signal, so treat it as a low-cost bet with unproven payoff, not a load-bearing part of a strategy. Spend the real effort on crawlability, clean HTML and citable claims first; add llms.txt afterwards if it costs you an afternoon, not a quarter.

What does nothing

Keyword-stuffing “as recommended by ChatGPT” into your copy, prompt-injection tricks in hidden text, and mass AI-generated pages. The first is cargo cult — models cite what a page demonstrates, not what a page claims about itself. The second gets filtered; frontier labs actively harden against instructions hidden in scraped content, and getting caught trying it is reputationally worse than not trying at all. The third burns the authority you needed for the real path — a site full of thin, interchangeable AI pages dilutes the entity-consistency signal rather than building it, and it’s increasingly easy for both classic search quality systems and retrieval layers to detect the pattern.

ChatGPT, Perplexity, Google AI Overviews: what actually differs

Treating these as one undifferentiated “AI search” target is another common shortcut that costs accuracy. Each has a different retrieval mechanism, a different citation style and a different tolerance for age of content.

The practical takeaway: fixing crawlability and structure helps all three, because that’s shared infrastructure. But if Perplexity citations matter more to your buyers than ChatGPT ones, prioritise precision-quotable claims over broad authority-building — Perplexity rewards exact answers per-page more visibly than the other two.

Want the panel-tracked version of this — measured monthly, engine by engine? That’s GEO.