B2B buyers start in ChatGPT. Your website still has to close them

G2 surveyed 1,076 B2B software buyers in March 2026. 51% now start research in an AI chatbot more often than in Google, up from 29%. The website is the close: proof, a parseable engagement shape, work without a login, and a next step that is not a 45-minute demo.

The first time a serious buyer hears your name, they are not on your homepage.

They asked ChatGPT which studio should rebuild the marketing site, or which agent builder fits a help centre that already exists. Three names came back. If you were one of them, they will open the URL the answer cited. They will stay if the page confirms the answer. They will leave if the page asks them to book a call before it will say what you do.

G2’s Tim Sanders, publishing The Answer Economy in April 2026, put it without decoration: you no longer buy the biggest ad or earn the top result. You win the answer. Then, if you are lucky, you get a visit from someone who already has a shortlist.

In the Answer Economy, Don't Win the Click — Win the Answer — G2, April 2026

LinkedInOpen the post

What did G2 actually measure?

The survey is 1,076 B2B software buyers and decision-makers, fielded in March 2026. It is software buying, not every B2B category. Treat it as the closest public read we have, not as a law of all commerce.

The numbers that change the homepage:

G2 also interviewed 39 B2B software marketers. Nearly all called AI discoverability a pipeline must-have. Most described themselves as early or still figuring it out. The lightbulb, more than once, was seeing a competitor in the answer and themselves missing.

The chatbot builds the shortlist. The website has to survive the second look.

We asked the same twelve buyer questions of ChatGPT, Perplexity, Gemini, and Google AI Mode and wrote down who got named. That log is four models, twelve questions. Use it as a method, not as a trophy.

The rest of search has been rehearsing the same compression. Ahrefs (December 2025 data) found AI Overviews correlate with a 58% lower CTR for position 1. Pew (July 2025) found an 8% versus 15% click on a traditional result when a summary is present, and a 1% click on a link inside the summary. SparkToro (June 2026) put U.S. zero-click Google search at 68.01% for January–April 2026. The buyer who still arrives is rarer, and further along.

Why is the website still the close?

Because the answer is a summary. The site is the evidence.

A chatbot will tell a buyer you run Meta and Google, or that you build agents. It will not — reliably — show the maison site, the night-line log, or the three-line note you send instead of a deck. Those things only exist if the page is public and specific.

G2 is blunt about trust. Buyers want proof. Citations from review sites were the top signal that increased confidence in a chatbot’s answer. Consistency across ChatGPT, Gemini, and Claude was itself a signal: the same vendor, described the same way. Inconsistency sent people looking for which version was true.

For a studio that is not a SaaS product, G2 Grid badges are not the whole job. The equivalent proof is work a stranger can open, a sentence about how you engage, and a next step that does not require a meeting to understand.

Designers have been feeling the other half of this: the site still has to hold attention, even as the first impression moves off-site.

A generated homepage that looks like every other generated homepage will confirm the buyer’s suspicion that the answer hallucinated you. The homepage test is four seconds: what does a stranger understand, and would they stay. That test gets stricter when the stranger arrived with a paragraph about you already in their head.

What must the page show?

Four things. If any one is missing, the visit dies in the same way — a back button, and a different name from the shortlist.

1. Proof that matches the answer

If ChatGPT said you build brand sites, the first case cannot be a stock dashboard. If it said you build agents, the first case cannot be a moodboard.

Our public work is four briefs, each with a constraint:

Sofia Rahman, founder of Eternal: “The site feels like walking into one of the rooms. Slow, dark, and exact. Collectors stay long enough to feel something.” That sentence is on the site because it is how the work was received. It is also the kind of quotation a model can lift without turning us into a generic web shop.

Proof is a named brief and a constraint, not a metric you would have to invent.

If the only proof is a logo row and “trusted by teams worldwide,” the buyer who came from ChatGPT has less than they already had in the chat. Show the work. Keep it readable without an account.

2. A parseable pricing and engagement shape

They will ask the chatbot what you cost. If the only public answer is “book a call,” the model will invent a package, or it will name a competitor who published a range.

We do not publish fake dollar amounts. We do publish the shape. Discovery is a written note on the work as it is. Build is the site, the campaign, the agent, the workflows — one team. Partner is the review after launch. The long version of that file is in llms.txt and the file stack. The human version of the money conversation is what a marketing website should cost.

Parseable means a person and a model can both answer:

• What are the shapes of the work?
• What is in each shape, and what is not?
• What do I send to start?
• What will I not get on the first call?

A calculator widget that hides the range behind six qualifying questions is not parseable. A PDF behind an email gate is not parseable. A sentence — “Discovery is a written brief; you keep it if we never work together” — is.

3. Work without a login

Retrieval cannot quote a case study that sits behind HubSpot. A buyer who already spent twenty minutes in ChatGPT will not create an account to see whether the studio can design.

Put the work on /work. Put the story in HTML, not in a slide embed that never renders for a crawler. Put the images on public paths. If a project is confidential, say so in one line and show a different one. Four honest cases beat twelve locked ones.

Google’s AI features guide is useful here even though this note is not about Google. They tell you to write non-commodity content, to organize it so a person can follow it, and to skip the tricks — no special markup, no llms.txt ranking claim, no page for every fan-out of a query. A public case with a point of view is the non-commodity object. A gated “insights hub” is not.

4. A next step that is not a 45-minute demo

The demo is what you offer when the page could not do the work.

A buyer who started in a chatbot has already compared you. They do not need a screen-share to learn that you build Next.js sites. They need a way to give you the current site, a week of tickets, or the ad account, and to get a written note back.

That is the close we run. Contact promises a note within 24 hours: what we would do first, and what we would not. The note is theirs to keep.

Bad next steps, in the order we see them:

• “Book a 45-minute discovery demo”
• “Talk to sales”
• A form with seventeen fields before a person will reply
• A chatbot that only captures an email and promises a recap

A chatbot on your own site can answer hours and locations. An agent can take the brief, file it, and only pull a person in when the work needs one. That is a later conversation. The first close is still a page that tells the truth and a next step a stranger can finish tonight.

What happens on the visit, in order?

The four-second homepage test still applies. If the first screen is a reel of adjectives, the buyer assumes the chatbot flattered you. If the first screen states the work — agents, campaigns, sites, from one studio — they will scroll for proof.

We would rather lose them on honesty than keep them on a reel. Eternal’s site is slow on purpose. A mid-market SaaS homepage that copies that pace for a product that needs a price and a login will fail the second look. Match the answer you wanted the model to give.

What should you stop asking the site to do?

Stop asking it to introduce you. The introduction already happened in the chat. The site confirms, or it contradicts.

Stop asking it to rank for “best AI agency” with a 3,000-word explainer that never names a case. That page is commodity. Google’s own examples of non-commodity content are first-hand and specific. A studio’s version is the brief, the constraint, and the Monday number.

Stop asking a calendar link to do the selling. G2 found 41% of buyers now run Deep Research — 10 to 20 pages of evaluation. Winning a footnote in that report is not a deal. Being one of the recommended names at the end is. Then those names get opened. A 45-minute demo scheduled three weeks out is how you lose a buyer who already did the research.

Stop hiding the engagement shape because legal has not approved a number. You can describe discovery, build, and partner without a rate. You cannot describe them if the only public noun is “solutions.”

What does this change on a marketing site this month?

Not a redesign. A pass over the pages a cited visit will actually open.

The homepage. Four seconds. Who you are, who you are not, what to open next. We wrote the test down separately.

The work index. Four cases, each with a brief a model can quote. No login. No “request the full case study.”

The services page. Marketing is inquiries, not reach. Agents are products, not experiments. Web is a site that can take the inquiry. Automation is the record moving between them. If a chatbot described you as a media buyer only, the marketing block has to correct that without a meeting.

The contact page. One specific ask. “Send the current site.” “Send a week of tickets.” “Send the last four weeks of landing pages.” We write back with what we would start and what we would leave.

The file the agents read. A short /llms.txt and a /pricing.md that state the three shapes. Not because Google ranks them — they do not — but because the same buyer who started in ChatGPT may ask a second model to “open the site and tell me how they engage.”

If you want a single picture of the studio at work, it is this: the site, the campaign, and the agent specified together, so the person who arrives from a chatbot lands on a page that can finish the conversation. Meridian was the campaign version of that. The weekly board is a decision, not a recap of screenshots. The site should be the same kind of object.

FAQ

Do B2B buyers really start in ChatGPT now?

For software, G2’s March 2026 survey of 1,076 buyers says 51% start in an AI chatbot more often than in Google, up from 29% in their 2025 report. 71% use a chatbot at some point, and 61% still use Google in tandem. That is a forked start, not a replacement. If your buyer is not buying software, treat the direction as real and the exact percentage as a software figure.

If the chatbot already named us, why does the website matter?

Because the visit is a check. G2 found 85% of buyers think more highly of a vendor the chatbot includes, and 69% switched their planned vendor on that recommendation. The site has to confirm the description, show proof, and offer a next step. If it contradicts the answer, or hides everything behind a demo, the buyer opens the second name on the list.

What should a B2B website show a buyer who came from ChatGPT?

Proof that matches the answer, a parseable engagement shape, work that does not require a login, and a next step they can finish without a 45-minute meeting. Those four objects are what a summary cannot carry. A logo row and a calendar link are what a summary already implied. If you have to choose, cut the reel and publish one honest case.

Is a 45-minute demo a bad first step?

As a first step, yes. The buyer who started in a chatbot has already compared you. A long demo before you will say how you engage, what a project looks like, or what to send, is a stall. Offer a written note in 24 hours against the current site or a week of tickets. Save the meeting for the work that actually needs one.

Should we gate case studies to capture leads?

No, not the cases you want cited. A gated PDF cannot be retrieved, cannot be quoted, and cannot confirm the chatbot’s description. Publish the brief and the constraint on /work. If a project is confidential, say so and show a different one. Lead capture belongs on the next step, after the proof has done its job.

How is this different from ordinary conversion rate work?

Ordinary conversion work assumes the homepage is the introduction. This assumes the introduction already happened in an answer engine. The page is asked to confirm, not to explain you from zero. That changes the first screen, the proof, and the call to action. It does not replace forms, speed, or clarity. It makes vague pages fail faster.

Do review sites still matter if buyers start in a chatbot?

Yes, especially later. G2 reports that review-site citations were the top trust signal for a chatbot’s answer, and that review sites gain influence as buyers move from consideration to decision. For a studio, the public equivalent is work, named quotes, and a consistent description across your own pages. Inconsistency across ChatGPT, Gemini, and Claude is a red flag buyers already act on.

What is the first page we should fix?

The URL the models already cite — or the URL you wish they would. Run the twelve buyer questions, write down the link each engine attaches to your name, and open it as a stranger. If the extract is missing, the proof is gated, or the next step is a demo, fix that page before you commission a new brand film.

Send the URL you think a buyer opens after ChatGPT names you, and the last version of the note you send after a first call. We will write back with what the page confirms, what it contradicts, and the next step we would put on it instead of a 45-minute demo. Contact the studio.