Your Buyers Are Building the Shortlist Inside a Chatbot. What a Solo Founder Can Actually Do About It.

AI chatbots now account for about 1.08% of referral traffic, per Conductor's 2026 benchmark. ChatGPT crawls roughly 1,500 pages to send one visitor. Googlebot crawls 18.

That's the traffic side. The shortlist side looks different.

What actually changed

G2 surveyed 1,076 B2B software buyers in March 2026 and found AI chatbots are now the top influence on vendor shortlists, at 54%. Review sites came second at 43%. Vendor websites, the thing most founders spend their weekends on, came third at 36%. Seventy-one percent of buyers now use chatbots for software research, up from 60%. Sixty-nine percent ended up choosing a different vendor than planned. A third bought from a company they'd never heard of.

A third of buyers discovering you for free is the best stat in B2B in a decade. It's also a third of buyers walking past someone who thought the relationship was locked. The reason traffic stays flat through all this is that the chatbot does the shortlisting first, so the visitor who does arrive is already half-sold. One qualified session instead of twelve exploratory ones. The funnel moved upstream instead of shrinking.

Why this suits one person better than SEO ever did

Evertune analyzed 200 million prompts and found the most-cited domain on any platform rarely clears 5% of citations. Wikipedia, Reddit, LinkedIn and YouTube combined rarely top 5% either. The rest is spread across thousands of domains nobody's heard of.

SEO rewarded whoever had spent the longest and the most, which is why a two-year-old company never really beat a fifteen-year-old one on a competitive term. Citation share is flat and long-tailed instead. Nobody owns the answer to "best inventory tool for a two-person ceramics studio" - it gets assembled fresh, from whatever's retrievable, every time someone asks. The competitive moat in AI search is specificity, which is exactly the asset a solo founder has more of than a Series C company does.

What actually moves the citation rate

The most-cited research is the Princeton GEO paper, which tested about 10,000 queries and measured what made a model more likely to quote a source. The top three tactics: adding statistics (a 41% lift), citing your sources (about 40%), and quoting a named expert (about 28%).

What they have in common is the opposite of how most founders write. "Teams love us" and "dramatically faster onboarding" were always weak copy, but a human reader used to skim past them to the screenshot. A retrieval system can't skim. It finds an extractable, attributable claim, or it moves to the next source.

The off-site half

Chatbots build 54% of shortlists, review sites build 43%, and your own site only 36% - most of your visibility is coming from places you don't own. Reddit is the extreme version: Peec AI's study of 30 million sources found it's the single most-cited domain overall, though the split by platform is wild - up to 47% of Perplexity's citations, about 21% of Google AI Overviews, single digits on ChatGPT, next to nothing on Gemini. Anyone selling a blanket "Reddit strategy" hasn't looked at the per-platform numbers.

The rest is duller: review-site profiles, category directories, and saying the same thing about yourself everywhere a model might read it. Models also carry stale memory - if you repositioned in the last eighteen months, some share of what's being said about you right now describes a company that no longer exists.

One popular fix is worth a caveat: publishing an llms.txt file. Adoption sits around 10% of domains, none of the major labs have committed to reading it in production, and in one 90-day crawl only 0.1% of AI bot visits targeted the file directly. It's a real move if you sell developer tools and buyers are pointing coding agents at your docs. For everyone else it's thirty minutes spent on the one item that feels like a technical fix instead of a writing problem, which is probably why it's so popular.

The case against caring about any of this

Steve Huffman told a 2025 earnings call that AI chatbots were "not a traffic driver" for Reddit, the most-cited domain on the internet. AI search handles somewhere between 12% and 18% of English-language informational queries as of Q1 2026, which leaves 82 to 88% happening elsewhere. And the conversion numbers people wave around, LLM-referred traffic converting far better than organic, carry an obvious selection effect: of course a visitor a machine personally recommended converts better. It's a small, highly qualified group of people arriving pre-sold.

I think that's right pre-revenue, and wrong the moment you have a category to be shortlisted in. The tell is whether buyers describe a comparison set when they talk to you. If someone opens with "we looked at you, Acme and one other," the shortlist already exists, and something is assembling it without you.

What this looks like with no budget

Assume a website, no agency, about four hours a month.

The diagnostic. The five questions a buyer would actually type, run through ChatGPT, Claude, Gemini and Perplexity, with the answers written down: whether you show up, who shows up instead, whether the description is accurate or eighteen months stale. Forty minutes, and it has to come first - the fix for "not mentioned" and the fix for "mentioned wrongly" are different jobs.

The pages and the claims. A self-contained answer at the top of five pages, with one number and one named entity in the opening paragraph. Every vague claim swapped for a checkable one - "faster onboarding" becomes "onboarding in 20 minutes, measured across 40 accounts in Q2 2026."

The entity record and the one surface. The same company name, category description and founding year everywhere a model might read them. One off-site surface, Reddit or your category's main review site, held for six months rather than five tried once. Plus one number from your own operations that nobody else can publish.

That's about fifteen hours across a quarter, no subscription required. What it costs instead is writing more specifically than founder marketing normally allows, because specific claims can be checked, and vague ones were never meant to be.

How you'd know it's working

You won't see it in analytics for a long while - 1.08% of referral traffic doesn't make a legible line on a chart. The measurable version instead is a monthly manual pass: the same five questions, the same four models, logged in a spreadsheet. Fifteen minutes, and it's the same method the paid tools (Profound, Evertune, Semrush One, Peec) run at scale, worth paying for once you have a position to defend rather than one to establish.

The unmeasurable version you'll notice first: buyers arriving already knowing what you do, with a comparison set they didn't get from you, asking a second-meeting question in the first meeting.

The bit that stayed with me

A third of B2B buyers now buy from a vendor they'd never heard of before a machine named it. Being unknown used to be a structural disadvantage that took money and time to fix, and the founders who won were disproportionately the ones who could afford to be seen. That's stopped being true, in favor of whoever's written down the most specific, most checkable version of what they actually do.

The work here is writing down exactly what your company is, with numbers attached, in language a machine can lift without hedging. Most founders haven't done that yet.