Most advice about getting your brand recommended by ChatGPT tells you to fix your website: add schema, write comparison pages, structure your headings as questions. That advice is not wrong. It is just aimed at the fourth-strongest signal.
When researchers measured what actually correlates with a B2B software brand being named in ChatGPT’s answers, three signals ranked above the brand’s own website — and all three sit on property you do not own. If your AI search strategy stops at your domain, you are optimising the weakest lever you have.
Why this matters now
The shift in where B2B buyers start has been fast enough to catch most revenue teams mid-stride.
- 51% of B2B software buyers now begin vendor research in an AI chatbot rather than a search engine — up from 29% eleven months earlier (G2, The Answer Economy: 2026 B2B Buyer Behavior Report, March 2026, n=1,076).
- AI-referred visitors convert at roughly 4.4x the rate of traditional organic visitors, because they arrive after the shortlist has been formed rather than before (Semrush, June 2025).
- 70% of marketers expect AI search to significantly change their strategy within three years. 20% have started (Acquia/Researchscape, 2025).
That last gap is the opportunity. The behaviour has already moved. The competitive response mostly has not.
Your Google rankings do not carry over
The most expensive assumption in AI search is that winning at SEO means winning at AI. Ahrefs tested it directly: 15,000 long-tail queries run through Google and Bing, then the same questions put to four AI assistants, with every citation mapped back against ranking position.
| AI assistant | Share of its citations that rank in Google’s top 10 |
|---|---|
| Perplexity | 28.6% |
| Gemini | 8.6% |
| Copilot | 8.2% |
| ChatGPT (in-text) | 8.0% |
| ChatGPT (references) | 6.1% |
Roughly nine in ten pages ChatGPT cites are not in Google’s top ten for the question that prompted them. Ranking first for your category term buys far less AI visibility than you would expect.
The mechanism is query fan-out. An assistant does not search your prompt. It decomposes the prompt into several sub-queries, retrieves against each, then synthesises. So the page that wins is not the one ranking for “best sales intelligence software” — it is the one ranking for whichever oblique sub-question the model happened to generate. You cannot reverse-engineer that reliably, which is exactly why signals travelling across all retrievals matter more than any single ranking.
What actually correlates with being recommended
A 2026 study analysed 3,508 citations drawn from 278 prompts across six B2B SaaS categories, tracking 95 brands and crawling 390 listicles and 141 Reddit threads to see which signals moved with how often ChatGPT named each brand.
| Signal | Correlation with brand mentions | Whose property is it? |
|---|---|---|
| Third-party listicle presence | +0.64 | Someone else’s |
| Branded web mentions | +0.58 | Someone else’s |
| Reddit presence | +0.57 | Someone else’s |
| Your own site being cited | +0.46 | Yours |
| G2 review count | +0.39 | Someone else’s |
| G2 star rating | −0.07 | Someone else’s |
Two rows deserve more attention than they usually get.
The first is the ordering. Your website is not irrelevant at +0.46 — but it is outranked by three signals you can only earn, never publish. The second is the last row. Star rating correlates at essentially zero, and slightly negative. Review volume moves with citations; review quality does not. A model reading review pages is extracting evidence that a product exists, is used, and is discussed. It is not grading you.
One caution before acting on any of this. These are correlations from a single snapshot, and the study’s own authors call it “small, directional” and warn it “should not be treated as universal ranking factors.” Brands with listicle placements also tend to be brands with marketing budgets, sales traction and PR. The honest reading is not “listicles cause citations.” It is that the brands models recommend are the brands the open web already discusses — and the open web is mostly not your website.
The four off-site signals, in priority order
1. Get into third-party listicles
Listicles are the single most-cited page type in the study at 18.8% of all citations, just ahead of vendors’ own product pages at 18.6%. When a buyer asks “what should I use for X,” the model is disproportionately reading somebody’s round-up post.
- Audit where competitors appear and you do not. Search your category’s “best tools” queries, list every round-up on pages one and two, and mark the ones naming three of your competitors while omitting you. That list is your outreach target, already ranked by relevance.
- Choose relevance over authority. A niche publication your buyers actually read outperforms a high-authority site in an adjacent category. The model is matching topical context, not domain prestige alone.
- Plan on following up. Around 80% of round-up placements are secured after a follow-up rather than on the first pitch. Treat it as a three-touch sequence, not a send.
- Write the entry for them. Supply a 60-word paragraph with the use case, the pricing tier and one honest limitation. Editors publish what is easy to publish, and a stated trade-off reads as credible rather than promotional.
2. Build co-occurrence, not just mentions
This is the finding almost nobody acts on. The average AI answer names 3.1 brands. Brands regularly appearing alongside six or more others averaged 9.1 mentions across the study. Brands appearing in isolation averaged 1.0.
Being named next to your competitors is not a threat. It is how a model learns you belong to the category at all. A page saying “these four tools all do X, and differ in Y” teaches a category membership that a page saying only “we do X” never will.
The same effect appears on your own pages: content naming six or more brands averaged 2.13 citations against 1.21 for content naming none. Comparison content that names rivals is not a concession. It is the highest-yield format you can publish.
3. Earn Reddit presence — carefully
Reddit’s weight is structural rather than accidental. Google reportedly pays around $60 million a year for Reddit data access and OpenAI an estimated $70 million, buying both training data and live retrieval. Perplexity draws 46.7% of its top-ten citations from Reddit alone.
It is also the signal most likely to burn you, in two separate ways.
Volatility. Reddit’s share of ChatGPT citations fell from roughly 60% to roughly 10% within two weeks in September 2025. Any channel a vendor can reweight overnight is one you rent, not own. Build presence there; do not build your plan on it.
Detection. Reddit communities remove promotional accounts efficiently, and a deleted thread is a deleted citation. The working convention is 90/10 — contribute genuinely useful answers roughly nine times for every one naming your product, and name it only where it directly answers the question asked. Accounts a few months old with minimal karma get filtered before a human reads them.
Target three to five category-specific subreddits rather than the large general ones. Expect three to four months before citation movement is measurable. If that timeline does not fit your quarter, this is the wrong signal to start with — begin with listicles.
4. Drive review volume, and ignore the average
Review count correlates at +0.39, star rating at −0.07. But the volume relationship is not linear, and its shape is genuinely surprising.
| G2 review count | Average mentions | Appearance rate in answers |
|---|---|---|
| None | 2.0 | 4% |
| 1–49 | 4.9 | 10% |
| 50–499 | 21.0 | 40% |
| 500+ | 10.6 | 27% |
Crossing fifty reviews is worth more per unit of effort than anything else on this list — a fourfold jump in appearance rate. Going from 500 to 1,000 appears to be worth nothing, most likely because the highest-volume brands are household names that models cite for reasons unrelated to their review page.
If you have fewer than fifty reviews, that is your highest-return project this quarter. Ask every closed-won customer. Do not filter for the happy ones — the rating does not appear to matter, and a mix reads as authentic to models and buyers alike.
What your own site still has to do
None of this means abandoning on-site work. It means sequencing it correctly, and knowing which parts are supported by evidence rather than folklore.
| What to do | Measured effect | Source |
|---|---|---|
| Add statistics to your claims | +41% AI visibility | Princeton / Georgia Tech / IIT Delhi, ACM KDD 2024 |
| Cite external sources in your content | +40% visibility | Same study |
| Front-load the answer | 44.2% of citations come from the first 30% of a page | SparkToro, January 2026 |
| Refresh existing pages | Median cited page is 3.9 months old; 69.7% published within 12 months | Position Digital, 2026 |
| Name entities densely | Cited pages average ~10 named entities per 1,000 words vs ~8 for uncited | Position Digital, 2026 |
The freshness figure is the one most teams underestimate. A median cited-page age of 3.9 months means AI search has a far shorter memory than Google. An evergreen page left untouched for two years is not evergreen in this channel — it is stale. Refreshing what already ranks beats publishing something new, and pages rebuilt from two-year-old originals reached a 38.6% citation rate.
One unglamorous prerequisite: confirm you are not blocking the crawlers. An audit of 50 SaaS sites found 68% were inadvertently blocking at least one major AI crawler through their WAF or robots.txt. Check that GPTBot, ClaudeBot, PerplexityBot and Google-Extended can reach you before spending a further pound on any of the above.
The platforms do not agree with each other
Treating “AI search” as one channel is the second-most-expensive assumption in this space. Across 680 million citations analysed in early 2026, only 11% of domains were cited by both ChatGPT and Perplexity.
- ChatGPT leans heavily on encyclopedic and editorial sources — Wikipedia alone accounts for 47.9% of its top citations.
- Perplexity leans on community discussion, drawing 46.7% of top-ten citations from Reddit, and pulls from three to four times more sources per answer than ChatGPT.
- Google AI Overviews sit closest to conventional search, with the highest overlap against organic rankings.
Being cited well in one engine tells you almost nothing about your position in another. Pick the engine your buyers actually use and measure that one, rather than averaging four dashboards into a number that describes nobody.
How to measure it without buying anything
Before subscribing to a tracking platform, run the manual baseline. It takes an afternoon and it is the only version you will fully trust.
- Write 30 buyer-intent prompts. Not brand searches — the questions a buyer asks before they know you exist. “Best tools for finding verified B2B contact data.” “How do I build a list of companies similar to my best customers.” Mirror how people actually type.
- Run all 30 in a logged-out session on each engine your buyers use. Logged-out matters: personalisation and memory will otherwise show you a flattering result nobody else sees.
- Record three things per prompt: whether you were named, which competitors were named, and which URLs were cited. That third column is your target list — it is a literal inventory of the pages influencing your category.
- Calculate mention rate. Named in 4 of 30 prompts is 13%. That single number is your baseline.
- Repeat monthly. Weekly is the recommended cadence for tracking tools, but monthly is enough to see whether off-site work is moving anything, and it is a cadence a real team will sustain.
The cited-URL column is the deliverable that matters. Every domain appearing there is a publication the models already trust for your category — which makes it the most precisely qualified outreach list you will build all year.
What this changes for the sales team
This is where AI search stops being a marketing project. If half of buyers now start in a chatbot, the shortlist is being drafted before anyone fills in a form, opens an email, or appears in your CRM as a lead.
Two consequences follow, and both land on sales rather than marketing.
Prospects arrive pre-loaded. The buyer on your first call has already read a synthesised comparison of you and three competitors, with no visible source and no opportunity for you to correct it. Discovery has to start by finding out what they were told, not by explaining what you do.
Absence is invisible. A buyer who never sees your name in an AI answer does not bounce off your website, does not abandon a form, and generates no signal anywhere in your funnel. You cannot report on a market you were never entered into. The only way to see it is to go and look — which is what the 30-prompt baseline above is for.
Off-site visibility and outbound prospecting solve the same problem from opposite ends. One gets you onto the shortlist before the buyer is in market. The other reaches the accounts where you were never on it. Teams that do only the second spend their days explaining who they are; teams that do both spend theirs explaining why they fit.
Where to start, by situation
| If this is true of you | Do this first | Time to signal |
|---|---|---|
| Fewer than 50 G2 reviews | Review drive across closed-won accounts | 4–8 weeks |
| Competitors in round-ups you are absent from | Listicle outreach, three-touch sequence | 4–12 weeks |
| Strong Google rankings, no AI mentions | Crawler audit, then refresh top pages | 2–6 weeks |
| Publishing a lot, cited rarely | Add statistics, cite sources, name competitors | 6–12 weeks |
| All of the above are handled | Reddit presence in 3–5 niche subreddits | 3–4 months |
Frequently asked questions
Is answer engine optimization just SEO with a new name?
No, and the overlap data proves it. Only about 12% of AI-cited URLs rank in Google’s top ten for the prompting query, falling to 8% for ChatGPT. The disciplines share techniques but reward different things: SEO rewards ranking for a query, AI search rewards being the source a model reaches for across many decomposed sub-queries.
How long before off-site work shows up in AI answers?
Listicle placements can register within weeks of publication, since retrieval indexes update continuously. Reddit presence typically takes three to four months of consistent contribution. Review volume moves in whatever time it takes you to ask your customers. None of it is instant, and anyone promising otherwise is selling something.
Should we pay for a listicle placement?
Paid placements on genuinely relevant publications behave much like earned ones for retrieval purposes. Paid placements on low-quality link farms do not, and carry conventional search risk on top. Judge by whether your buyers would plausibly read the publication, not by its domain rating.
Does a bad review hurt our AI visibility?
On the available evidence, no. Star rating correlated at −0.07 with brand mentions, which is statistically indistinguishable from zero. Volume of discussion is what moves. This may change as models get better at weighing sentiment, but today a mixed review page beats an empty one comfortably.
Can we just ask employees to prompt ChatGPT about us?
Some agencies sell this as “engagement campaigns.” There is no credible public evidence that individual prompting influences what a model recommends to other users, and the mechanism most people assume is at work does not exist in the way described. Treat it as unproven and spend the budget on reviews.
Which engine should we optimise for first?
Whichever one your buyers use, which you find out by asking them during discovery rather than by inference. If you have no data at all, start with ChatGPT on reach, and note that its citation profile favours editorial and encyclopedic sources over community discussion.
The takeaway
The instinct to fix your own website first is understandable, because it is the only surface you fully control. It is also why so many AEO programmes stall: they pour effort into the one signal the data ranks fourth, then conclude the channel does not work.
Models recommend brands the open web already discusses. Round-ups, mentions, communities and reviews are that discussion. Your website is your account of yourself, and a system built to weigh evidence treats it accordingly.
Run the 30-prompt baseline this week. If you are named in fewer than one in five answers, the work is off-site, and the ordering above tells you where to start.
Request a demo and we will run a live search against your best-fit account profile. Or explore the Sales Bundle, and read more in our B2B Growth Hacks topic hub.