The sales call is going well until the rep asks the routine question — “how did you hear about us?” — and the prospect says the sentence that’s been showing up in more discovery calls every quarter: “I asked ChatGPT for the best options and you were on the list.”

Wonderful. Except the marketer listening to the recording afterward has a follow-up nobody can answer: why were we on the list? Which prompt? What did it say about us? Are we on the list for the queries that matter, or just that one? And the uncomfortable inverse — for how many prospects did ChatGPT build a shortlist we weren’t on, closing the deal against us before our name entered the room?

That’s the shape of the newest channel in B2B: OpenAI confirmed ChatGPT passed 900 million weekly users in early 2026, processing roughly 2.5 billion prompts a day, and the company’s own usage research found about half of that activity is people seeking information and recommendations — the queries that used to start on Google now ending in a single synthesized answer with a handful of named brands. Unlike rankings, there’s no results page to audit. But the recommendation isn’t random, and it isn’t unreachable. ChatGPT decides which brands to name through mechanisms you can understand, influence, and measure — and this guide covers all three.

How Does ChatGPT Actually Decide Which Brands to Recommend?

Through two distinct knowledge sources, and your playbook depends on knowing which one is answering.

Source one: trained knowledge. For many recommendation prompts, ChatGPT answers from what its models learned during training — the accumulated public web, where your brand exists as an entity with associations: what you do, who you serve, how you’re described across sites, reviews, communities, and comparisons. This source updates slowly (model releases, not news cycles), can’t be directly edited, and rewards one thing above all: consistent, corroborated presence over time. If the public web describes you fifty slightly different ways, the model’s internal representation of you is fuzzy — and fuzzy entities don’t get confidently recommended.

Source two: live search. When a prompt benefits from current information — “best X in 2026,” pricing, comparisons — ChatGPT searches the live web, retrieves pages, and composes a cited answer. This source runs on crawler eligibility: OpenAI’s bots (GPTBot for training, OAI-SearchBot for search) must be able to access and read your pages — and critically, they don’t execute JavaScript, so content rendered client-side is invisible to them. From the retrieved candidates, selection follows the extraction logic every serious citation study keeps confirming: direct answers near the top of the page, verifiable specifics, clean structure, and sources whose claims corroborate across the web.

Layered on both: ChatGPT’s shopping and recommendation surfaces (product cards and structured suggestions, which OpenAI has stated are organic rather than paid), and memory-based personalization that tailors answers per user — meaning two people asking the same question can get different lists, which has measurement implications we’ll get to.

The strategic summary: trained knowledge decides whether you’re a candidate; live retrieval decides whether you’re cited today; and both run on the same fuel — a consistent entity attached to extractable, specific, corroborated content.

The 6-Step Playbook

Step 1: Establish your baseline — before touching anything

Define 20–30 prompts your buyers plausibly ask — category recommendations (“best [category] for [audience]”), comparisons, problem-phrased queries (“how do I solve X”) — and record what ChatGPT answers today: are you named, in what position and framing, and who appears instead? Run each prompt fresh (memory personalizes results, so use clean sessions for baseline work) and repeat on a schedule, because answers shift with model updates and retrieval changes. This is tedious manually and continuous by design — it’s exactly what Search Ranking Intelligence automates, tracking your brand across ChatGPT alongside Claude, Gemini, Perplexity, Grok, and Google, so the baseline becomes a trend line instead of a quarterly chore.

Step 2: Verify crawler eligibility — the ten-minute gate

Check that your robots.txt doesn’t block GPTBot or OAI-SearchBot (some sites blocked AI crawlers wholesale in 2023–24 and forgot), confirm your key pages render their substance in HTML rather than client-side JavaScript, and make sure your most recommendable pages — comparisons, pricing, use-case pages — are indexed and fast. Uncrawlable content can’t be cited, full stop; this step is boring and non-negotiable, and it’s a developer conversation, not a content one.

Step 3: Fix your entity — the highest-leverage unglamorous work

One canonical description of what you are, who you serve, and how you differ — enforced everywhere: homepage, about page, directories, review profiles, LinkedIn, everywhere your brand appears. Entity fuzziness is the silent disqualifier in trained-knowledge answers, and consistency is the fix that compounds across every future model release. This is the Knowledge Base’s structural job in Iriscale: your positioning, ICP, and approved terminology as one source of truth applied to everything the platform produces, so the consistency survives busy quarters and team changes.

Step 4: Publish the answers ChatGPT wants to lift

For each priority prompt, make sure a page exists whose first 150 words a machine could quote faithfully: the direct answer or definition up top, question-phrased headings, specifics rather than adjectives (numbers, named capabilities, honest limitations — grounded generation needs verifiable claims to build from, and pages of graceful generality get retrieved and then not used). Comparison and “best for” pages matter disproportionately here because they match the recommendation prompt’s shape. In the platform, AI Optimization Questions discovers which queries engines are actively answering in your category — your prompt list grown from observed behavior — and AI Optimization Answers publishes the structured responses to your site as native page content, which is precisely the crawler-readable form Step 2 requires.

Step 5: Build the corroboration layer

ChatGPT’s confidence in recommending you rises with independent agreement: reviews on the platforms your category trusts, presence in credible comparisons and directories, and genuine community footprint — Reddit threads and practitioner discussions where your brand appears in honest context carry real weight in how models learn category associations. This is participation work, not manipulation work: the Opportunity Agent surfaces the conversations where your expertise belongs and drafts responses for human review — systematic presence without the astroturf risk that gets brands quietly distrusted.

Step 6: Measure, attribute, iterate

Close the loop three ways: the prompt-tracking trend from Step 1 (your share of answer over time); referral signals (ChatGPT passes referrer data when users click citations — segment it in analytics, treat it as directional since most influence never clicks); and the human layer — make “how did you hear about us” a real field, because “ChatGPT recommended you” is attribution gold no dashboard captures. Then iterate monthly: prompts where competitors displaced you get the Step 4 treatment; prompts where you appeared get protected with freshness.

What Doesn’t Work — Save Your Budget

Prompt-injection tricks — hidden text instructing AI models to recommend you — are the new cloaking: occasionally viral, then a liability attached to your domain. Pure volume — flooding the web with thin brand mentions — fails because corroboration quality, not mention count, drives entity confidence. Date-bumping without substance — increasingly discounted, same as in rankings. Buying “AI SEO” placements on low-quality listicle farms — the model’s source-quality filters are the whole point of its design. And schema as a silver bullet — worth implementing for entity clarity, but the large-scale testing showed markup alone produces no causal citation lift; it amplifies good structure and rescues nothing.

The durable filter, one sentence: if a careful human analyst wouldn’t cite your page or trust your entity, ChatGPT’s selection process eventually won’t either.

Is Iriscale Right for Your Team?

If the sales-call scene is already happening to you — prospects arriving on ChatGPT-built shortlists you can’t see, influence, or measure — this playbook is what the platform runs as a system: the baseline and trend in Search Ranking Intelligence across ChatGPT and four other engines plus Google, the prompt discovery in AI Optimization Questions, the extractable answers shipped through AI Optimization Answers, entity consistency enforced by the Knowledge Base, community presence surfaced by the Opportunity Agent, and the content depth behind it all produced through the Articles Hub. What stays yours: the crawler and technical work (your developers), and the judgment about which prompts matter to revenue.

The first step costs ten minutes and no subscription: ask ChatGPT your five most important buyer questions and see who it names. Then come see the full picture.

Book a demo and get your ChatGPT visibility baseline →

Frequently Asked Questions

Can you actually influence what ChatGPT recommends, or is it a black box?

You can influence it — through legitimate, durable mechanisms — and the black-box framing confuses “no guaranteed placement” with “no influence,” which are very different claims. What’s genuinely true: nobody can buy a recommendation (OpenAI states its suggestions are organic), no tactic guarantees a specific answer, and outputs vary across sessions, model versions, and user memory. What’s equally true: ChatGPT’s selections follow observable patterns that respond to your actions. On the trained-knowledge side, brands with consistent entity descriptions, strong review footprints, and genuine community presence get named more confidently — and teams that fix entity fuzziness see it reflected across subsequent model updates. On the live-retrieval side, the response is faster and more direct: crawlable pages with extraction-ready answers to the exact prompts buyers ask demonstrably enter citations within weeks, because retrieval rewards the best liftable answer available right now. The honest expectation-setting: influence operates on probabilities, not switches — you’re raising your selection odds across thousands of prompt variations, which is why measurement (tracking your share of answer over time) matters more here than in any channel you’ve run before. Black boxes can’t be measured into strategies; this one can, which is the entire premise of treating it as a channel.

How long does it take to start appearing in ChatGPT’s answers?

Two clocks, matching the two knowledge sources — and knowing which clock governs which work prevents both false hope and premature surrender. The fast clock is live retrieval: when ChatGPT searches the web for current-information prompts, a newly published or restructured page that’s crawlable and carries the best extractable answer can appear in citations within days to weeks. This is why the playbook front-loads the answer-page work — it’s the fastest external validation available, and teams frequently see their first cited appearance before their first ranking movement on the same content. The slow clock is trained knowledge: the entity associations that drive confident unprompted recommendations (“best tools for X” answered without live search) accumulate across the public web and consolidate at model-update cadence — months, not weeks — which is why entity consistency and corroboration work should start immediately even though its payoff lags. The compounding pattern most teams experience: early citation wins on retrieval-served prompts (weeks one through eight), gradually broadening presence as the corroboration layer builds (months two through six), and the trained-knowledge payoff — being named by default — emerging across subsequent model releases. Baseline before starting, or none of this is provable; and judge the program at ninety days on trend direction, not on owning every prompt.

Why does ChatGPT recommend our competitors but not us?

Run the diagnosis in order, because the four causes have four different fixes and guessing wastes quarters. Cause one, eligibility: your pages may be invisible to OpenAI’s crawlers — robots.txt blocks from the 2023–24 AI-blocking wave, or substance rendered in client-side JavaScript that GPTBot and OAI-SearchBot can’t read. Ten-minute check, developer fix. Cause two, extraction: you’re crawlable but your answers are buried — the competitor’s tidy comparison table and first-paragraph definition get lifted while your equivalent insight sits in paragraph nine of a 3,000-word post. The fix is the answer-page restructuring in Step 4, and it’s the most common cause we see. Cause three, entity fuzziness: the model can’t confidently resolve who you are — inconsistent descriptions across your site and profiles, a name shared with other entities, positioning that changed twice — so it recommends the competitor it can resolve. Fix: canonical description, enforced everywhere, patiently. Cause four, corroboration gap: the competitor genuinely has the stronger independent footprint — more reviews, more community presence, more third-party comparisons — and the model’s confidence follows the evidence. Fix: the participation work of Step 5, which is slower and completely legitimate. The diagnostic that sorts them: check whether ChatGPT knows you at all (ask it directly about your brand). Known-but-not-recommended points to extraction or corroboration; unknown-or-confused points to eligibility or entity. Each is fixable; none is fixed by publishing more of what you already publish.

Should we block GPTBot to protect our content, or allow it for visibility?

This is a genuine trade-off with a clear answer for most B2B brands — allow it — and a framework for the cases where it’s genuinely contested. What blocking protects: your content from training future models (GPTBot) and from appearing in ChatGPT’s search citations (OAI-SearchBot, if blocked). What blocking costs: exactly the visibility this guide exists to build — blocked content can’t be cited, and a brand absent from training data accumulates entity fuzziness while competitors accumulate associations. For a B2B SaaS company whose content exists to be found by buyers, the arithmetic is lopsided: your comparison pages, guides, and answers are marketing assets whose entire purpose is discovery, and blocking the fastest-growing discovery surface to protect them is protecting the ad from the audience. The genuinely contested cases: publishers whose content is the product (paywalled journalism, proprietary research), where training-data extraction competes with the business model — there, granular control makes sense: many split the decision, blocking GPTBot (training) while allowing OAI-SearchBot (citations with referral traffic), which captures visibility while limiting training use. The audit worth running today regardless of your position: check what your robots.txt currently says, because a meaningful number of sites blocked everything during the 2023–24 wave, forgot, and are now invisible to the channel their own marketing team is trying to win — the single most common “why aren’t we cited” answer that takes ten minutes to find and one line to fix.

Do ChatGPT recommendations actually drive revenue, or is this vanity visibility?

The influence is real and mostly dark — which makes it a measurement design problem, not a vanity question. Three evidence layers, honestly weighted. First, behavioral scale: with roughly half of 2.5 billion daily prompts seeking information and recommendations per OpenAI’s own usage research, the surface where shortlists form has provably moved — the only open question is whether your category’s buyers are there, which your own baseline answers. Second, direct referrals: ChatGPT passes referrer data on citation clicks, and teams tracking it report modest but high-intent traffic — treat it as the visible tip, because the dominant pattern is influence-without-click: the buyer reads the answer, forms the shortlist, and later types your name directly, which analytics records as brand search or direct traffic. Third — the layer that settles it internally: self-reported attribution. Teams that add “AI assistant recommendation” to their how-did-you-hear fields and train sales to log the discovery-call answer consistently find it appearing within a quarter, and growing. The reporting frame that keeps this honest: track share of answer (your presence across the tracked prompt set) as the channel metric, direct referrals as a floor, and self-reported attribution as the revenue connector — and resist both errors: dismissing the channel because clicks look small, and claiming every brand-search uptick for it. The sales-call test remains the most persuasive artifact in any budget conversation: when your own recordings contain “ChatGPT recommended you,” the vanity debate tends to end.

Does ChatGPT’s memory and personalization mean everyone sees different recommendations?

Partially yes — and it changes how you measure, more than how you optimize. ChatGPT’s memory features let it tailor answers to what it knows about a user: their role, stack, past conversations, and stated preferences can all shape which brands surface and how they’re framed. Two users asking “best [category] tool” may get overlapping-but-different lists — one skewed toward their company size, another toward tools that integrate with software they’ve mentioned. The measurement implication is immediate: baseline and tracking work must use clean, memory-free sessions for consistency (which is how systematic tracking operates), while accepting that real buyers’ answers will vary around that baseline — your tracked share-of-answer is the center of a distribution, not a universal fact. The optimization implication is more interesting and mostly favorable: personalization rewards segment clarity. A brand whose public footprint says precisely who it’s for — “built for B2B SaaS teams of one to fifty” — gives the model the material to surface it for exactly those users, while vague every-company positioning gets averaged out of everyone’s personalized answers. This aligns the channel with what good positioning demanded anyway: the sharper your ICP claim across your entity footprint, the more often personalization works for you, putting you in front of the buyers you’d have chosen — which is arguably the first discovery channel where narrowing your claimed audience widens your effective reach.

Is optimizing specifically for ChatGPT worth it, or should we optimize for all AI engines at once?

Do the shared work once, then let measurement tell you where engine-specific attention pays — the either/or framing dissolves under the mechanics. The large majority of what wins ChatGPT recommendations — crawlable pages, extraction-ready answers, entity consistency, corroboration footprint — is precisely what wins Claude, Gemini, Perplexity, and Grok, because all five run variants of the same selection logic: retrieve, resolve, extract, ground. That shared core is 80 percent of the program and it’s engine-agnostic; running it “for ChatGPT” versus “for AI search” is the same work with different labels. Where engine-specific reality enters: the citation studies consistently show low overlap in which sources each engine favors — different corroboration diets, different retrieval freshness, different index coverage — so your results will differ by engine even when your inputs don’t, and ChatGPT-specific eligibility (GPTBot/OAI-SearchBot access) has exact parallels per engine worth checking individually. The operating model that follows: one optimization program, five scoreboards — which is exactly why Search Ranking Intelligence tracks all five engines plus Google in one view, and why single-engine tracking misleads in whichever direction flatters. ChatGPT earns priority attention for one defensible reason: scale — 900 million weekly users makes it most categories’ largest answer surface — but the teams that win the era are optimizing the shared core and letting per-engine measurement direct the marginal 20 percent, rather than running five separate playbooks or pretending one engine is the whole game.

What should a small team do first if we can only do one thing this month?

Ship one genuinely excellent answer page for your single most valuable recommendation prompt — because it exercises the entire playbook in miniature and produces the fastest provable win. The sequence, one focused week: pick the prompt (the “best [category] for [your ICP]” question closest to your revenue), baseline it (ask ChatGPT in a clean session, screenshot who’s named — your before), then build or restructure the page that deserves to be cited: the direct answer in the first 150 words, question-phrased headings, a comparison table if the prompt implies one, honest specifics including who you’re not for (models demonstrably favor sources that acknowledge limits), and your canonical entity description verbatim. Verify eligibility the same day — robots.txt allows OpenAI’s crawlers, content renders in HTML — and link the page from your strongest existing pages so it isn’t an orphan. Then re-check the prompt weekly and log what changes. Why this beats every alternative first move: it’s fully within one person’s control, it front-loads the fast clock (retrieval-served prompts respond in weeks), it forces the entity and extraction disciplines you’ll scale later, and — most practically — it produces the artifact that unlocks everything else: a before-and-after screenshot pair showing your brand entering an answer it was absent from. Nothing in this channel funds the bigger program — the tracking, the corroboration work, the platform conversation — like one demonstrated win, and nothing demonstrates it faster than the single-prompt sprint.

Related Reading

• How to Implement Generative Engine Optimization
• AI Search Optimization vs Traditional SEO
• How to Embed AI Answers Into Web Pages
• Cross-Engine Visibility Share: The Content ROI KPI
• How to Evaluate AI Content Optimization Success

© 2026 Iriscale · iriscale.com · AI-Powered Growth Marketing for B2B SaaS