How B2B SaaS Buyers Actually Use AI to Build Shortlists: The 2026 Evidence

Summary: Forrester's 2026 Buyers' Journey Survey of roughly 18,000 global business buyers found that 94% now use generative AI somewhere in the purchase process, up from 89% in 2025. Gartner's separate 2026 study of 645 buyers put genAI usage during a recent purchase at 45%. Both numbers are correct. They measure different things — and the gap between them is the single most useful fact in this article.

The shortlist forms before you know the deal exists

For most of the last two decades, a B2B software deal had a legible beginning. Someone searched a category term, landed on a comparison page or a review site, downloaded something, and entered a CRM. Marketing could see the shape of demand because demand travelled through instrumented channels.

That beginning has moved. A buyer now opens ChatGPT or Gemini and types something closer to how they actually think: "we're a 40-person team on Postgres, what should we use for product analytics that won't cost a fortune." The engine returns three or four names, a paragraph of reasoning, and a handful of citations. By the time anything shows up in your analytics, the consideration set is already closed.

The evidence for how widespread this has become is unusually good for a shift this recent, because two large research organisations measured it independently.

Forrester's 2026 Buyers' Journey Survey, covering nearly 18,000 global business buyers, found generative AI use in the purchase process rose from 89% in 2025 to 94% in 2026. Forrester also reported that twice as many buyers named generative AI or conversational search as their most meaningful research source than named any other single source — ahead of vendor websites, product experts, and sales representatives.

Gartner's 2026 study of 645 buyers found 45% used generative AI during a recent purchase, mostly to gather information on vendors and products.

Ninety-four percent and forty-five percent are not contradictory. Forrester asked whether AI appeared anywhere in the process across a very large multinational sample. Gartner asked about a specific recent purchase in a smaller sample. One measures penetration; the other measures per-deal incidence.

The practical read: near-universal exposure, roughly half of individual deals materially touched. Either number is high enough that "we'll wait and see" is no longer a defensible position — but you should know which one someone is quoting at you before you build a budget on it.

What buyers ask AI to do

Aggregated 2026 buyer research puts the most common AI tasks in vendor evaluation at:

That third row deserves more attention than it usually gets. A buyer building a business case is asking the engine to argue on your behalf to their CFO — using whatever description of your product the model has assembled from third-party sources. If that description is two pricing tiers out of date, or attributes a competitor's compliance certification to you, or omits the integration that justifies the purchase, the business case is built on it anyway.

This is a different failure mode from being absent. Being absent costs you the deal. Being inaccurately present can cost you the deal later, more expensively, after procurement has already been briefed on something untrue. Detecting and disputing those descriptions is what Entity Intelligence exists to do.

Gartner has also projected that 90% of B2B buying will be agent-intermediated by 2028 — a forecast, not an observation, and worth treating as directional rather than a planning number. But it points the same way as the measured data.

The uncomfortable part: nobody agrees on what gets cited

Here is where practitioners should slow down, because this is where most GEO advice quietly falls apart.

A reasonable question is: if we rank well on Google, do we get cited by AI? Several credible organisations have measured this and produced answers that cannot all be right.

A spread from 12% to 76% is not a rounding difference. It reflects genuinely different methodologies: different engines, different query mixes, different definitions of "cited," different sampling windows. Semrush's comparison work found that Google's own AI Overviews and AI Mode share only 13.7% URL overlap with each other — two surfaces from the same company citing largely different sources.

Two conclusions follow, and they matter more than any individual number.

First: your Google rankings are not a proxy for your AI visibility. They may correlate strongly or barely at all, and which one is true for your category is an empirical question, not something you can reason your way to. We go deeper on where the two disciplines overlap and diverge in AEO vs SEO.

Second: be sceptical of any vendor — including us — quoting a single confident number about AI citation behaviour without publishing how they got it. This is why CiteRank replays every prompt 10–30 times per engine rather than querying once, and reports a 95% confidence band on every score. Single-shot prompting produces numbers that look precise and aren't. The full protocol, including sampling, is published in our methodology.

Which sources the engines actually lean on

If Google position isn't the reliable lever, what is? Two large-scale citation analyses give partial answers:

• Peec AI analysed roughly 30 million citations (March 2026) and found Reddit the most-cited domain across ChatGPT, Google AI Mode, Gemini, Perplexity and AI Overviews combined.
• Profound analysed 1.4 million citations (November 2025 – February 2026) and found LinkedIn the most-cited domain for professional queries across six major AI platforms.

For a SaaS marketer this is an awkward finding, because neither domain is one you control. You cannot publish your way onto Reddit's front page, and you can't schema-mark a forum thread.

What you can do is treat the citation supply chain as a mappable asset rather than a mystery. For any prompt cluster in your category, a finite set of domains supplies the engine's answer. Some are earnable — analyst pages, comparison sites, documentation aggregators, community threads where your users already are. Working out which specific domains feed the prompts you lose is the job of citation intelligence, and it produces a materially different content plan than keyword research does.

Note what this rules out. It rules out the assumption that publishing more pages on your own domain is sufficient. It often isn't, because the engine may not be reading your domain at all when it answers the prompt that matters.

The five prompt clusters that decide SaaS deals

Not all prompts are worth equal effort. In B2B software, five clusters carry most of the commercial weight, and they behave differently.

1. Category discovery — "best product analytics tool for a Series B startup." Answers rarely name more than three or four products. Highest stakes, hardest to enter, most defensible once won.

2. Alternatives and displacement — "alternatives to [incumbent]." Structurally easier for challengers, because the engine is explicitly looking for a list of non-incumbents. Usually the fastest first win.

3. Head-to-head comparison — "X vs Y." If a competitor has published the comparison page and you haven't, the engine reads their framing of you. This is the cluster where silence is most expensive.

4. Integration and ecosystem — "best CRM for a team already on HubSpot." Decided almost entirely by how machine-readable your integration documentation is. Frequently the highest-conversion cluster and the most neglected.

5. Compliance and security — "is [product] SOC 2 compliant." When nothing structured and authoritative exists, engines are prone to fabricating an answer. A hallucinated negative here can quietly kill enterprise deals with no trace in your funnel.

The right unit of analysis is the prompt, not the keyword. A keyword tells you what people typed; a prompt tells you what question the engine is answering and which stage of the buying process it belongs to. Weighting prompts by buying-stage value — and knowing which ones you lose — is the core of the AI Recommendation Share metric.

A defensible measurement approach

If you're setting this up internally, the sequence that survives scrutiny looks like this:

1. Build the prompt graph, not a keyword list. Cover all five clusters above, tagged by intent, sized to your category. Most B2B categories are reasonably covered by 250–750 prompts.
2. Replay, don't sample once. LLM outputs are non-deterministic. A single query is an anecdote. Ten to thirty replays per prompt per engine gives you a distribution you can put a confidence interval around.
3. Track every engine your buyers actually use, not just ChatGPT. ChatGPT dominates referral volume, but Gemini, Perplexity and Copilot show meaningfully different citation behaviour — and Copilot matters disproportionately in Microsoft-heavy enterprises.
4. Separate absence from misrepresentation. "Not mentioned" and "mentioned inaccurately" need different fixes and different urgency.
5. Attribute changes to specific interventions. Ship one change class at a time and re-run, or you'll never know which fix moved the number. This is where revenue attribution earns its keep — connecting a citation gain to pipeline rather than to a vanity score.

You can run all of this manually. It's tedious rather than difficult. The reason teams eventually automate it is the replay volume: 500 prompts × 8 engines × 20 replays is 80,000 API calls per cycle, and you need it weekly for the trend line to mean anything. The mechanics are covered in AI visibility tracking.

What to do this quarter

If you take one thing from the data above, take this: the 12%-to-76% disagreement about citation overlap means your category's behaviour is unknown until you measure it. Not unknowable — unknown. That's a solvable problem, and it's cheap to solve relative to the deals moving through it.

A sensible first pass:

• Write down the ten prompts a real buyer in your category would type. Not keywords — full sentences, with the constraints a real buyer would mention.
• Run each one across ChatGPT, Gemini, Perplexity and Copilot. Three times each, minimum.
• Record: are you named? In what position? Described accurately? Which sources did the engine cite?
• Look at what the cited sources have in common. That's your target list.

That exercise takes an afternoon and will tell you more than any general GEO article, this one included. If you'd rather see the full version — 250 prompts across all eight engines with per-engine breakdowns — you can run a free AI visibility audit, or look at a labelled sample report first to see exactly what the deliverable contains.

Related reading

• What Is Generative Engine Optimization (GEO)? The Complete 2026 Guide — the foundational concepts behind everything above
• Why ChatGPT Recommends Another Clinic Instead of Yours — the scoring model and its limits
• AI Visibility for B2B SaaS — the vertical playbook
• 98.8% of Multi-Location Brands Are Invisible to ChatGPT — how selectivity plays out in local search
• GEO for Financial Services — the same problem under regulatory constraint

Sources

All figures in this article come from third-party published research. No CiteRank client data appears here.

• Forrester, 2026 Buyers' Journey Survey (~18,000 global business buyers)
• Gartner, 2026 B2B buyer study (645 buyers); Gartner agent-intermediation forecast for 2028
• Ahrefs, AI Overviews citation overlap analysis (2025)
• BrightEdge, AI citation overlap analysis
• Search Atlas, LLM citation study (18,000+ queries)
• Semrush, AI Mode vs AI Overviews comparison study
• Peec AI, citation analysis of ~30 million citations (March 2026)
• Profound, citation analysis of 1.4 million citations (November 2025 – February 2026)

Published 7 August 2026. Figures reflect the most recent published versions of each study as of that date; AI search research is revised frequently and readers should check for updates.