ChatGPT Drives 92% of AI Referral Traffic. Here's the B2B SaaS Read

Where do AI answer engines actually send B2B SaaS buyers, and which engine matters most? As of mid-2026, the answer is lopsided: ChatGPT drives about 92% of all trackable AI referral traffic. That figure comes from Previsible's 2026 State of AI Discovery Report, the third in its AI Traffic Study series, which analyzed 6.77 million LLM-referred sessions across 166 GA4 properties from November 2024 through May 2026. It is the largest standalone-LLM referral study Previsible has published, and the story it tells is consolidation.

The headline is easy to repeat and easy to misread. Below is what the data says, the one caveat that reframes it, and the finding underneath the headline that actually changes what B2B SaaS teams should build. For the groundwork on how answer engines differ from classic search in the first place, start with our guide to AEO versus SEO for B2B SaaS.

What the data actually says

Four numbers carry the report.

ChatGPT commands 92.4% of trackable LLM referral traffic and is still gaining share. At the end of 2025 the same measurement put ChatGPT around 84%, with Perplexity near 9% and Gemini near 4.5%. Since then ChatGPT has pulled further ahead while most challengers lost ground. It is the only standalone LLM sending meaningful referral volume across most industries and page types.

Total AI referral volume grew 9.9x in 19 months. Monthly LLM-referred sessions across the sample rose from roughly 65,000 in November 2024 to 644,000 in May 2026. The growth was not smooth. There was a sharp single-month drop in late 2025 tied almost entirely to a ChatGPT change, then a recovery. The direction is clear even if the month-to-month is noisy: search is moving toward the chat experience, and the chat experience is capturing more of it.

The challengers the industry bet on did not win. Perplexity peaked in March 2025 and has fallen 61% since. Copilot collapsed 96% from its August 2025 peak. The workplace-embedded-AI thesis, the idea that Microsoft would win discovery through Office, did not hold in the referral data.

Gemini is the quiet number two, and Claude is the fast riser. Gemini grew steadily to take the second spot, though its real footprint is likely undercounted because so much of it lives inside Google Workspace and Android rather than trackable web referrals. Claude grew 64x over the window and overtook Perplexity in March 2026. More on Claude below, because it is the platform most relevant to a slice of the B2B SaaS market.

Optimizing for a generic "AI visibility" without prioritizing ChatGPT is optimizing for an abstraction. One engine sends the traffic.

The caveat that changes the headline

Here is the line most people repeating the 92% number skip: the study deliberately excludes Google AI Overviews.

Overviews and AI Mode live inside Google's own results. They do not generate trackable referral sessions the way a click from chatgpt.com does, so they operate on a different measurement paradigm and sit outside this dataset. The report is candid that AI discovery happening inside Google almost certainly represents more volume than every standalone LLM platform combined.

So the accurate way to state the finding is narrow and specific: within trackable standalone LLM referral traffic, ChatGPT owns 92%. That is a real, large, growing slice. It is not all of AI search. If you treat the 92% as the whole picture, you will underweight the AI discovery that happens where you cannot see a referrer at all, which is exactly why a single AI-visibility number is misleading. This is the same blind spot we walk through in the guide to measuring AI visibility: no one surface sees the whole of it, so you triangulate GA4, Search Console, and direct citation tracking rather than trusting any one of them.

The finding that matters most for B2B SaaS

Strip away the market-share horse race and the most useful finding in the report, for a B2B SaaS company, is about where AI traffic lands, not which engine sends it.

For SaaS specifically, the largest landing surface for LLM-referred traffic is the internal search page. The engine knows your domain answers the question. It cannot confidently pick which of your pages does. So instead of citing a specific URL, it routes the visitor to your site and effectively hands them your search box. The report frames this as a structural pattern of the search-style engines, ChatGPT and Gemini: they trust the domain, then get uncertain at the page level. It shows up across verticals and time periods, which suggests it reflects how retrieval-augmented generation works rather than a temporary quirk.

That reframes internal search from a navigation convenience into an acquisition surface. If an engine sends a high-intent visitor to your search results page and your search returns weak results, you lose a user the AI specifically chose to send you. Sites with strong search UX convert those sessions. Sites without one leak them at the front door.

The deeper move is to stop the handoff from happening at all. The engine defaults to your search box when it cannot find a single obviously-correct page to name. You reduce that ambiguity three ways:

• Give every buyer query its own page. When the answer to "best X for Y" is a dedicated, well-titled URL rather than a filter on your search results, the engine has something specific to cite. This is the whole case for comparison pages: they hand the engine a page built to be the answer.
• Make the right page easy to reach and rank. A page that is well linked internally gets crawled, indexed, and pulled into the retrieval pool across more of the sub-queries an engine fans a question into. An orphaned page often is not in the index at all, so it cannot be picked. The mechanics are in our piece on internal linking for SaaS clusters.
• Resolve every page to one clean URL. When the same content is reachable through parameters, subdomains, or trailing-slash variants, the engine can hedge across versions or cite a non-preferred one. Consolidating those signals is covered in the guide to fixing SaaS canonical conflicts.

Domain trust is the hard part, and if an engine is already sending you search-page traffic, you have it. The page-picking problem is the fixable part. Most B2B SaaS teams are sitting on earned domain trust and losing the visit at the last step because the engine could not find one page worth naming.

Claude stopped being a rounding error

Claude grew 64x over the study window and overtook Perplexity in monthly referral sessions in March 2026. The absolute volume is still small next to ChatGPT, but two things make it worth attention for a subset of B2B SaaS.

First, the referrals are hard-won and high-intent. Second, Claude behaves differently from the search-style engines. It is a content-selection model: it picks specific pages and skews toward long-form, research-oriented content such as guides and documentation, and its referred users engage with that content at higher rates than any other platform's. Its growth tracks its expansion into coding tools, professional workflows, and enterprise adoption.

That profile maps directly onto a lot of the B2B SaaS market we work in. If your buyers are developers, technical operators, or professional-services teams, which describes a large share of PropTech, FinTech, and ConstructTech audiences, Claude visibility is becoming a real factor rather than a speculative one. It rewards exactly the depth of content that the search-style engines are too uncertain to cite at the page level, which means the same investment in thorough, well-structured pages pays off in two places at once. This lines up with what we found running the same query set across engines in our PropTech citation research: the engines do not agree, and a page built for depth wins the selection-style engines even when the volume leader ignores it.

The PropTech and FinTech read

Aggregate numbers flatten the part that matters for a vertical SaaS company. Two vertical findings in the report are worth pulling out.

SMB software carried one of the highest AI penetration rates of any vertical. Penetration here means the share of a vertical's total sessions that come from LLMs. SMB-focused software led the field. For a PropTech or FinTech platform selling into small and mid-sized operators, that means AI discovery is not a someday channel. It is already a larger fraction of your traffic than the site-wide average across the web would suggest.

Financial services showed some of the highest AI density on conversion pages. In finance and insurance, blog content pulled the largest share of LLM traffic, as you would expect. But the conversion pages, the signup, enrollment, and quote-request surfaces, showed some of the highest AI-to-total ratios in the entire dataset. AI is not just sending finance buyers to read. It is sending them to the pages where they act.

Put those together and the implication for a FinTech or PropTech SaaS is direct: AI discovery is already reaching your highest-intent pages, not only your top-of-funnel content. The pricing page, the product page, and the signup flow are AI entry points now. That raises the stakes on making those pages legible to engines, which for most teams means transparent, machine-readable pricing instead of a "contact us" wall that gives an engine nothing to summarize or compare.

What to actually do now

Five moves, in priority order, grounded in the data above.

1. Build your AI plan around ChatGPT first. It is the one engine sending meaningful referral volume. Get the fundamentals of citation right there before spreading effort thin across platforms that send a trickle. Our practical playbook for getting cited in ChatGPT is the place to start.

2. Fix the page-picking problem. Audit where your AI traffic lands. If it is piling onto your internal search page, the engine is telling you it trusts you but cannot find your best page. Give it dedicated pages for buyer queries, link them well, and resolve URL variants.

3. Measure by page type, not site-wide. Your overall AI penetration rate might be a fraction of a percent while your pricing page runs several times higher. A site-wide average hides where the traffic concentrates and where the leverage is. Segment your GA4 view by template.

4. Make pricing machine-readable. If AI is reaching your conversion pages, "contact sales for pricing" leaves the engine with nothing to compare or recommend. Clear, structured pricing gives it something to cite.

5. Monitor Claude if your buyers are technical. You do not need to reprioritize the whole plan, but if you sell to developers or professional-services teams, add Claude to your citation tracking now. Early positioning in a growing selection-style engine compounds.

The takeaway

The consolidation headline is real and worth acting on: ChatGPT is where trackable AI referral traffic lives, and an AI plan that does not lead with it is misallocated. But the number to internalize is not 92%. It is the quieter finding underneath, that engines trust B2B SaaS domains and then stall at the page level, dropping high-intent visitors on internal search pages because no single page was obviously the answer.

That is a good problem to have, because domain trust is the hard half and you already earned it. The fixable half is giving the engine one clean, well-linked, purpose-built page to name. Do that on the pages your buyers actually convert on, measure it by page type, and keep an eye on Claude if your audience is technical. The engine already wants to send you the visitor. Your job is to make sure it knows which door to point them at.

Frequently asked questions