Half Your Software Buyers Never Visit Your Site

More than half of software decision-makers now begin their research inside an LLM rather than a search engine. Sixty-nine percent report choosing a different vendor than the one they originally planned to buy, based on what an AI told them. Roughly a third bought from a vendor they had never heard of before the AI introduced it into the conversation.

Read that last figure again, because it is the one that should change your marketing plan, not the first two. A third of your category's buyers are being introduced to vendors they did not know existed, by a system that decides which names get spoken and which do not. If your name is not one of them, you are not losing a deal you were in the running for. You never entered the room.

What "Research" Actually Means Now

The traditional B2B software buying journey has a well-worn shape: a pain point triggers a search, the search produces a shortlist, the shortlist gets narrowed through demos and reviews, and a decision emerges from a committee. That shape has not disappeared, but a new step has inserted itself at the very front of it, before search, sometimes before the buyer has even fully articulated the problem to themselves.

A decision-maker with a vague sense that their current tool is not working types a conversational description of the problem into ChatGPT or Perplexity, not a keyword. The AI's response is not a list of ten links to evaluate. It is two or three names, with reasoning attached, functioning as a pre-formed shortlist the buyer did not have to build themselves. Everything that happens afterward, the demos, the committee discussion, the vendor comparison, happens inside the frame that initial answer set.

The Buying Committee Still Exists

It is tempting to read the LLM-first statistic as evidence that B2B software buying has become an individual, impulsive decision. It has not. Enterprise software purchases are still made by committees, with procurement, IT, finance and the actual end users all weighing in at different stages, and that structure has not meaningfully changed.

What has changed is which committee member does the earliest reading, and what they bring back to the group. Increasingly, that first pass is an AI-generated shortlist rather than an analyst report or a Google search session, and the committee's subsequent evaluation starts from whatever names made that initial cut. A vendor absent from the AI's answer is not evaluated and rejected. It is simply never raised.

Ask It Two Ways, Get Two Different Kinds of Answer

Not every prompt behaves the same way, and understanding the difference matters more than most GEO advice acknowledges. A generic category prompt, "what is the best CRM," tends to retrieve the same aggregator-driven answer search already gives you: a ranked list drawn from review platforms. A constraint-rich prompt, "which CRM works for two hundred salespeople, integrates with Salesforce for finance, and operates in regulated financial services," behaves completely differently. It tends to surface vendor documentation, integration pages, and niche analyst coverage instead of the generic top-ten list.

Prompt Type	Example	What Gets Retrieved
Generic category	"What is the best CRM?"	G2, Capterra, top-10 listicles
Constraint-rich	"CRM for 200 salespeople, Salesforce integration, regulated finance"	Vendor docs, integration pages, niche analysts

This has not been formally studied for B2B software specifically, and should be treated as an observed pattern rather than an established law. But it is consistent with how these systems are known to behave more broadly, and it has a direct strategic implication: content built to answer the specific, constrained questions your actual buyers ask has a real chance of being retrieved. Content built to rank for the generic category term is competing directly with the aggregators, in a fight most vendors cannot win.

Do AI Engines Even Agree on How Direct to Be?

Different engines behave differently when asked for a software recommendation, and this is worth knowing before you build a monitoring programme around a single platform. Perplexity tends to give ranked, sourced, comparative answers, the closest thing to a direct recommendation among the major engines. ChatGPT without live search access tends to hedge, disclaiming that it cannot verify current pricing or reviews, while ChatGPT with search behaves much more like Perplexity. Gemini often defers to official product pages and professional advisors for enterprise decisions rather than naming a clear winner.

No systematic study has measured refusal or hedging rates across engines for software queries specifically, so treat this as a working observation rather than a statistic to cite externally. The practical takeaway is that a monitoring set covering only one engine will give a distorted picture of your actual visibility, because the engines are not behaviourally interchangeable.

The Committee Member Nobody Is Marketing To

Most SaaS demand generation is still built around the assumption that a human will encounter an ad, a landing page, or a piece of content, and click through. When the first touchpoint is an AI-generated answer summarising your product to someone who never visits a page you built, the entire funnel model that demand generation reporting is built on stops describing what is actually happening.

This is not a reason to abandon demand generation. It is a reason to stop treating clicks as the only signal that counts. A buyer who reads an AI summary of your product, forms an opinion, and later arrives directly at your pricing page having skipped every touchpoint your analytics were built to track, still went through a research journey. It is simply one your current reporting cannot see.

What Actually Changes in a Marketing Plan Because of This

The tactical response is not to abandon SEO or paid acquisition, both still matter for the buyers who do search directly. It is to add a second, parallel workstream aimed specifically at being named inside AI-generated answers: structured comparison content, accurate and complete review platform profiles, documentation built for citation, and a monitoring programme that checks what ChatGPT, Perplexity and Gemini actually say about your category on a regular cadence.

The monitoring piece deserves particular attention because it is the part most companies skip entirely. Without a defined, consistently-run prompt set, you have no way of knowing whether you are gaining or losing ground in this channel, and anecdotal spot-checks produce misleading confidence in either direction. A fixed set of ten to fifteen prompts, representative of how real buyers actually ask, run monthly across the major engines, is a small operational commitment that produces the only reliable signal available right now.

What Gets Lost When the Buyer Skips the Website Entirely

There is a specific kind of persuasion that only happens on a vendor's own site: the carefully sequenced case study, the interactive demo, the pricing page built to anchor a comparison in your favour. When an AI summarises your product instead of a buyer reading that sequence directly, all of that sequencing disappears, replaced by whatever the AI decided was worth mentioning, in whatever order it decided to mention it.

This is a real loss, and no amount of GEO optimisation fully recovers it. What GEO optimisation can do is influence which facts and framing make it into that AI summary in the first place, by making sure the content an engine retrieves already contains the comparison points, the differentiators and the honest caveats you would want a human salesperson to raise. You are not writing for a page anymore. You are writing raw material for someone else's summary.

Why the Companies That Panic Tend to Get This Wrong

The most common reaction to these numbers, once a marketing team actually internalises them, is to try to produce a large volume of AI-facing content quickly. This usually underperforms a much smaller, slower effort, for a reason that keeps surfacing throughout GEO research: retrieval systems weight credibility and structure far more heavily than volume, and a sudden spike of thin, similar-sounding content reads as exactly the kind of scaled content production these systems are increasingly tuned to discount.

A smaller set of genuinely well-researched, honestly-written comparison and evaluation pages, published over a longer window and maintained rather than abandoned after launch, consistently outperforms the volume approach.

What a Sales Team Notices Before Marketing Does

Sales reps often sense this shift before any dashboard confirms it, because they are the ones hearing prospects open a call already knowing surprisingly specific things about the product, or asking pointed questions that came from somewhere other than the website. Treating those anecdotes as a formal input to content planning, rather than dismissing them as anecdotal, is a low-cost way to catch this shift early in a specific category before the aggregate statistics confirm what is already happening deal by deal.

The Committee Roles Nobody Maps Onto AI Behaviour Yet

Procurement, IT security, finance and the end-user champion do not disappear from the buying committee just because an AI shortened the front end of the process. What changes is which role is most exposed to AI-generated information first, and it is rarely procurement. The end-user champion, the person who will actually use the software daily and often initiates the search for a replacement tool, is typically the one typing a conversational problem description into ChatGPT before anyone else on the committee is even aware a purchase is being considered.

This has a specific practical consequence: the AI-generated shortlist frequently becomes the champion's opening pitch to the rest of the committee, arriving already framed by whatever the AI decided to emphasise. A vendor absent from that first answer is not just missing from a search result, it is missing from the internal pitch the champion brings to procurement and finance. Winning the champion's individual research moment is now doing work that used to happen later, in a formal RFP or vendor comparison document.

What Happens to the RFP When the Shortlist Arrives Pre-Formed

Formal RFP processes assume a roughly open field: procurement issues requirements, multiple vendors respond, a scoring rubric picks a winner. When three names arrive pre-selected by an AI before the RFP is even drafted, the RFP quietly becomes a formality that ratifies a decision already shaped upstream, rather than the mechanism that makes the decision.

This is not a hypothetical concern raised for effect. It follows directly from the finding that a third of buyers purchase from a vendor they had never heard of before an AI introduced it, combined with the reality that most B2B RFPs are written with at least one informal favourite already in mind. If that informal favourite increasingly comes from an AI-generated shortlist rather than word of mouth or a previous vendor relationship, the RFP's actual function shifts without anyone formally deciding it should.

None of this is easily reducible to a single ROI figure, and pretending otherwise sets up a credibility problem later. The honest framing is that this is an awareness and consideration investment with indirect revenue effects, measured through citation frequency trends against a fixed prompt set, branded search movement, and post-conversion survey data asking buyers how they first heard about you. Our book Cited or Silent covers the measurement and ROI question in far more depth, including the RoGEO framework this reporting approach is built on. This shift in where research happens sits directly upstream of the category-term problem covered in our piece on SEO for SaaS.

Why the Same Content Performs Differently Across the Buying Committee

A technical comparison page written for the end-user champion and a compliance-focused page written for procurement are not interchangeable, and an AI answer often blends signals from both without distinguishing which committee member actually needs which framing. This matters for content strategy in a way most GEO advice glosses over: a single comparison page trying to serve every committee role simultaneously tends to serve none of them particularly well, while separate pages addressing the champion's day-to-day usability question and procurement's risk and compliance question each have a cleaner shot at being retrieved for the specific prompt that role is actually likely to ask.

This is a genuine trade-off against the "fewer, better pages" advice given elsewhere in this content series. The resolution is not a contradiction: fewer, better pages still beats a flood of thin content, but "fewer" does not have to mean one page trying to answer every committee role's question at once. Two or three well-targeted pages, each clearly aimed at a specific role's actual question, is consistent with both pieces of advice at the same time.

---

Frequently Asked Questions

Does this mean SEO no longer matters for SaaS?

No. It means a second, parallel workstream aimed at AI citation now matters alongside SEO, not instead of it. Buyers who search directly still exist and still convert.

Has the B2B buying committee structure actually changed?

Not structurally. What changed is which committee member does the earliest reading and what they bring back, increasingly an AI-generated shortlist rather than an analyst report or search session.

Do all AI engines recommend software the same way?

No. Perplexity tends toward direct, sourced comparisons. ChatGPT hedges more without live search access. Gemini often defers to official pages for enterprise decisions.

How do we monitor this without expensive tooling?

Build a fixed set of ten to fifteen prompts representative of real buyer questions, and run them monthly across ChatGPT, Perplexity and Gemini.

What should we tell leadership about ROI?

Frame it honestly as an awareness and consideration investment with indirect effects, measured through citation frequency and branded search trends, not a single attribution figure.

Why do constraint-rich prompts matter for content strategy?

They appear to retrieve a different, less aggregator-dominated set of sources than generic category prompts. Content built to answer specific buyer constraints has a realistic chance of being cited where generic content does not.

What is the biggest mistake companies make reacting to this shift?

Treating it as a reason to panic-produce generic AI content at volume. Retrieval systems reward credibility and structure, not volume, and a flood of thin content tends to underperform a small set of well-structured, honest comparison pages.

---