There is no reliable list of "AI ranking factors" the way there's a rough industry consensus on SEO ranking factors built from two decades of testing. Nobody outside the AI labs knows the exact mechanism a given model uses to decide which brand to name in a given answer, and that mechanism can shift with every model update. Anyone selling you a guaranteed way to "hack" ChatGPT into recommending you is overselling — the tactics below are the ones grounded in how these systems are known to actually work, and they're about improving your odds, not guaranteeing an outcome.
First, let ChatGPT’s search crawler in
ChatGPT answers from two places: what the model learned in training, and a live search when the question needs current information. OpenAI runs a separate crawler for each job and says “each setting is independent of the others” (OpenAI crawler documentation):
User agent
What OpenAI says it does
Top sites blocking it
OAI-SearchBot
Surfaces websites in ChatGPT’s search features. Opted-out sites won’t appear in ChatGPT search answers, though they can still show as navigational links
6.0%
ChatGPT-User
Fetches pages for user-initiated actions; robots.txt rules may not apply
7.8%
GPTBot
Crawls content that may be used to train foundation models
14.8%
Share of 4,762 robots.txt files from the Tranco top 10,000 that disallow each crawler from the homepage, read October 28, 2026. Method in our crawler study.
OpenAI recommends allowing OAI-SearchBot in robots.txt and from its published IP ranges, and says a change takes about 24 hours to apply. Most top sites already get the split right: of the 705 that block GPTBot, 421 (60%) still allow OAI-SearchBot. But 100 of the 284 sites that block OAI-SearchBot never name it; a catch-all User-agent: * rule does it for them. Check yours before any of the advice below, because a blocked search crawler means you can’t be cited from search however good the page is.
Be unambiguous about what you are
AI assistants — whether pulling from training data or a live web search step — do better with brands that are easy to categorize. If your website, your reviews, and your third-party mentions all consistently describe you the same way ("expense management for mid-market finance teams," not "fintech platform" on one page and "spend solution" on another), a model has an easier time confidently associating you with the exact question a buyer asked. Category ambiguity doesn't just confuse search engines — it gives a language model less to work with when it's deciding who to name.
Earn real third-party mentions
This is the single most consistently cited factor across the research available. Ahrefs' analysis of AI visibility across roughly 75,000 brands found that mentions on YouTube had the strongest correlation with AI visibility of any signal tested, and branded web mentions overall (press, review sites, comparison content) also correlated strongly — see Ahrefs' brand visibility correlation study. Separately, reporting on how ChatGPT surfaces recommendations points to third-party validation — press coverage, analyst mentions, and consensus across high-authority sources like Wikipedia — as a bigger factor than on-page SEO tricks (Profound). Practically: get covered by real publications in your space, get listed in real comparison/roundup articles, and don't manufacture fake ones — a model summarizing thin or spammy sources tends to reflect that thinness right back.
Be genuinely well-reviewed
Review sites such as G2 and Capterra are among the pages assistants read about software. If your average rating is mediocre or your reviews are inconsistent across platforms, that's a real, visible signal a model can pick up on when it's assembling a recommendation — not a rumor. This isn't something you can shortcut convincingly; it's downstream of actually being good at what you do and asking real customers to say so.
Publish content that directly answers buyer questions
Write for the actual question, not the keyword. "What's the best X for Y" is close to how a real buyer phrases it to an AI assistant — content that answers that framing directly, in plain language, with a clear point of view, is easier for a model to extract and quote than content built around keyword density. Structure helps too: clear headers, direct answers near the top, and content a model can lift a sentence from without needing to infer meaning across paragraphs.
Keep your public facts consistent and current
Wrong or outdated information about your business anywhere prominent online — an old Google Business Profile listing, a stale "About" page, conflicting details across directories — can get folded into what a model believes about you, and there's no way to "correct" a model after the fact except by fixing the source. Accuracy is unglamorous, but it's one of the few genuinely controllable levers here.
What doesn't work (and why)
Keyword-stuffing your own site with your brand name doesn't move a model's decision to recommend you — models are synthesizing meaning, not counting keyword frequency the way an old search algorithm might.
Buying fake reviews or fabricated "mentions" risks the opposite of the intended effect if it's ever surfaced or contradicted elsewhere online, and it's the kind of thing that erodes the actual trust signals these systems are trying to detect.
Prompt injection or trying to manipulate a specific model's output directly isn't a real GEO strategy — it might work once, in a demo, and it doesn't scale to real buyers asking real questions in the wild.
Which fix goes with which gap
The tactics above work better when you know which gap you're closing. These are the patterns worth looking for in the answers, and the fix that usually goes with each:
A competitor is quoted where you aren't. Read the sentence the assistant uses about them and pull out the concrete claim: "easiest to set up", "best for pipeline reporting". Publish the same kind of fact about yourself, plainly, and get it onto the review and comparison pages that cover your category.
One buyer question never names you. If you're named for "best CRM for small sales teams" but never for "switching from spreadsheets", the gap is the framing. Cover that question in its own words on a page you control.
One assistant never names you. Check whether it searches the web before answering or answers from training data. A search-grounded model needs you on pages it can fetch and cite; a training-data model changes slowly, through broad coverage over time.
The assistant cites sites you're not on. When a model returns its sources, those are the review sites, docs and threads to get covered on.
How to know if any of this is working
The only reliable way to know is to ask the models the same questions your buyers ask and read what comes back, not to infer it from search rank. Mentioned generates 5 realistic buyer questions for your category, asks Claude, ChatGPT and Gemini each one independently, and shows the sentence where you're mentioned, or who got recommended instead when you're not.
To see whether a change worked, track the keyword rather than re-running one-off checks. A one-off check writes fresh questions each time, so comparing two of them mixes your change with a change in the questions. A tracked keyword's questions are locked and asked every week, so a move after your change is a real move. On Starter and above, the Actions screen ranks the gaps above by how many answers each one is costing you, with a playbook of the usual fixes under each.
Rule out access first: that ChatGPT, Claude and Perplexity can read your site at all. A blocked search crawler means you can’t be cited however good the page is. AI crawlers and robots.txt explains which bots matter, and the AI crawler checker tests yours.
