What a B2B Website Actually Needs in 2026, Now That Half Your Traffic Never Clicks

Two things changed at once: AI summaries absorb the click your page used to get, and page speed became a measurable revenue variable rather than a technical nicety. A B2B site built for 2026 answers the buyer's question completely on the page itself, loads fast enough that the answer arrives, and is structured so an AI assistant can extract and attribute it. The old model — a brochure site that ranks and waits for a form fill — is the one losing pipeline.

Published July 22, 2026 · Last updated July 22, 2026

What actually changed, with numbers

The click got absorbed. When an AI summary appears above Google's results, users click a traditional result on 8% of visits, versus 15% when no summary is shown, and only 1% click a link inside the summary itself, according to a study of 68,879 real searches by 900 U.S. adults from the Pew Research Center (2025). An AI Overview is a machine-generated summary that sits above the organic results and answers the query on the results page, which is why it removes the reason to visit the source.

The loss concentrates where you spent the most to get. AI Overviews reduce the click-through rate of the number-one position by roughly 34.5%, based on an analysis of 300,000 keywords by Ahrefs (2025). Ranking first still beats ranking fifth; it is worth about a third less than it was.

The compensation has not arrived. AI referral traffic accounted for 0.14% of total traffic in 2025, growing 66% but from a very small base, while organic search still accounted for 16.04%, per a study across more than 50,000 sites by Semrush (2026). That ratio is the planning problem: the damage is here and the replacement channel is roughly a hundred times smaller.

For a B2B company, fewer people arrive and they arrive later, having already read a summary of you written by something else. The page has to close a decision, not start one.

The 9 pages that still generate pipeline

Nine page types still earn their keep, and most B2B sites have three of them. Each answers a question a buyer asks out loud, specifically enough that an assistant can quote it.

The problem page, not the product page

A problem page is a page that describes the situation the buyer is in, in their words, and only then explains what you do about it. Buyers search the symptom before they know the category name. Name it in the heading and let the product arrive last.

Pricing or pricing logic

If you cannot publish numbers, publish the model: what drives cost up, what drives it down, what a typical engagement includes, what is billed separately. A buyer who leaves to find your price elsewhere reads someone else's version of it.

Proof: a real case with numbers

One case study with a starting figure, an ending figure, a time frame and a named constraint outperforms twelve logos. Include what did not work. Assistants quote specifics and skip adjectives.

Integration pages

One page per system you connect to — CRM, ERP, helpdesk, billing — stating what the integration reads, what it writes, whether it is native or via API, and where the limits are. They catch buyers checking compatibility with a stack they already chose.

Comparison pages

Your buyer compares you whether or not you host the page. Write the honest version: where the alternative wins, where you win, and what switching costs. A comparison that concedes nothing reads as marketing.

The question pages your buyer asks before the call

Implementation time, who does what, what happens to existing data, what the contract looks like, what happens if it fails. These consume the first two calls. Answering them in public shortens the cycle.

A short, qualifying contact form

Six fields, one qualifying question, and a clear statement of what happens next. If part of your market operates in another language, route it properly instead of translating the confirmation email — the failure that breaks machine-translated support breaks forms too, the argument for bilingual AI agents in the intake path.

Team and credibility

Real names, real roles, real locations, and the legal entity behind the brand. This is what an assistant reads to decide whether you are a company or a landing page.

A genuinely useful tool or calculator

A calculator, a benchmark, a template that works without an email address. It is the only asset here that earns links and gets named by other people.

How do you write a page an AI assistant will cite?

By making one passage on the page quotable on its own. Answer engine optimization is the practice of structuring a page so an assistant can extract a specific passage, attribute it, and reproduce it without distortion. It sits on top of ordinary SEO: the page still has to be crawlable, indexed and ranked first. Use this template:

1. Write the H1 as the full question a buyer would type into a chat, not a keyword fragment.
2. Put a 40 to 60 word direct answer in the first paragraph, before any context or setup.
3. Make every H2 a question that stands alone, and answer it in the first paragraph below it.
4. Keep sections between 150 and 350 words, so a retrieved fragment is complete on its own.
5. Never open a section with "as we saw above": the reader may arrive at that section alone.
6. Put comparisons in a real HTML table with a header row, not in a paragraph or an image.
7. Attach a named source, a year and a link to every number, in the same sentence as the number.
8. Add FAQPage schema whose answer text matches the visible text, and show the published date.

Two things decide whether any of it matters. Check that robots.txt does not block GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot or Google-Extended: a blocked crawler cannot cite you. And make sure the answer exists in the HTML source, not only after a JavaScript render.

Speed is now a revenue line item

Speed is a revenue variable because a page that arrives late does not get read, and the loss shows up at the form. Core Web Vitals are three field metrics that measure loading, responsiveness and visual stability, assessed at the 75th percentile of real user data over 28 days. The thresholds are LCP at or under 2.5 seconds, INP at or under 200 milliseconds, and CLS at or under 0.1 — and only 48% of sites pass all three on mobile, according to the Web Almanac 2025 from HTTP Archive and the definitions published on web.dev.

The conversion side has one figure worth knowing. A 0.1 second improvement in mobile speed produced a 21.6% increase in form submissions on lead generation sites, in a Deloitte study for Google (2019) covering 37 brands and more than 30 million sessions. Treat it as directional: it is from 2019, the sample skews toward large consumer brands, and your own baseline decides how much room there is.

Frame it as a budget line. Take your monthly sessions, your form conversion rate and your average deal value, then model what a tenth of a second is worth on your own numbers. If you also buy traffic, the delay charges you twice, because the slow page wastes the click you already paid for — one of the reasons why your Google Ads cost per lead keeps rising. Most B2B sites lose their speed to a heavy page builder, uncompressed images, five tracking scripts and a font stack nobody audited.

The pricing page argument

A buyer who cannot find your pricing does one of two things: leaves for a competitor who published theirs, or fills in the form purely to extract a number, which produces a call with someone who has no budget. Both cost you, and only the second shows up as a lead.

Publishing a range is not the same as publishing a rate card. What a buyer needs before a call is the order of magnitude and the variables that move it. A page that gives the model — scope tiers, what each includes, what is billed separately, a typical timeline — qualifies as well as a number and commits you to less.

When you genuinely cannot publish a figure, say why. DINOLABS ranges in U.S. dollars for a B2B site are [RANGE TO BE CONFIRMED]: our published pricing is set in Colombian pesos and the USD conversion has not been validated, so we describe what each tier contains instead of quoting a figure we cannot stand behind. "It depends" followed by the three things it depends on is an answer; "contact us for pricing" is not.

Forms: qualify, don't collect

The job of a B2B form is to separate buyers from browsers, not to maximize submissions. Six fields do it: name, work email, company, role, what they are trying to solve, and a budget or timeline question with ranges as options. Everything else you can enrich or ask on the call.

The budget question is the one most companies delete, and it is the one that filters. Ranges rather than a blank box make it answerable without negotiation, and a prospect who picks a range far below your floor has saved both sides a meeting. Pair it with a statement of what happens next: who replies, within how long, and what the first call covers.

Two mechanics matter more than the field count. Show validation errors inline instead of on submit, and never clear a filled form on error. And measure the form as a funnel: views, starts, completions and the field where people abandon.

What should you measure when the click no longer happens?

Measure the visibility you keep and the outcome you close, because the middle of the funnel is now partly invisible. Four indicators cover it without inventing a metric nobody acts on.

Impressions vs clicks in Search Console as an AI Overview signal

Compare impressions and clicks per query over time. A query where impressions hold or rise while clicks fall and average position is unchanged is the signature of an answer served above you. Build that list first, then decide which pages to rewrite for extraction.

Referrals from chatgpt.com, perplexity.ai and gemini

Segment referral traffic from assistant domains and track it as its own channel. The volume will be small — 0.14% of total traffic across the Semrush sample cited above — but those sessions arrive further along in the decision, so judge the channel on meetings booked, not sessions.

Branded search volume as a downstream indicator

When an assistant summarizes your category and names you, the visible effect is often a search for your brand rather than a referral. Track branded query volume in Search Console monthly. A rise with no campaign behind it is the closest proxy for being mentioned in answers you cannot see.

Form-to-meeting rate instead of raw traffic

Traffic is a weak proxy for demand when fewer visits can accompany the same pipeline. Report submissions, qualified submissions and meetings held, plus the rate between each step. If traffic drops 20% and meetings hold, your pages are doing their job.

What not to do

Five common responses to AI search make the situation worse. They are worth naming because each one is sold as a solution.

Publishing volumes of AI-generated pages with no data behind them is the first. It produces text that says what every other page says, which is what a summarizer discards, and it dilutes the pages that were working. Gating top-of-funnel content is the second: a PDF behind a form cannot be read, quoted or attributed, so it adds nothing to the visibility you are defending. Keep gates for proprietary material.

Chat widgets that cover the page or delay it are the third — they compete with your own answer and damage the metrics you are measuring. Carousels are the fourth: they push your best message off the screen and rarely survive contact with mobile field data. The fifth is using "AI-powered" as the value proposition. It describes your implementation, not the buyer's outcome, and every vendor in every category now claims it.

A realistic redesign sequence

Fix speed and the three highest-intent pages first, then redesign. Most companies do the opposite: a six-month redesign, a launch, and a performance audit afterwards, which is when the traffic loss shows up. The other order lets you learn from real behavior first.

Phase one, weeks one and two: measure current field data, fix images, scripts and fonts, and get Core Web Vitals to threshold on the templates that already receive traffic. Phase two, weeks three through six: rewrite the pricing page, the top problem page and the main comparison page for extraction, and rebuild the form as a qualifier. Phase three: everything else, informed by what the first two produced. DINOLABS ranges in U.S. dollars for this work are [RANGE TO BE CONFIRMED] while the conversion from our published Colombian peso pricing is validated.

Copy this checklist into your redesign plan:

• ☐ Record current LCP, INP and CLS from field data, not from a lab test.
• ☐ Export the 12-month impressions-vs-clicks curve per query from Search Console.
• ☐ Confirm robots.txt allows GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot and Google-Extended.
• ☐ Verify the answer text exists in the HTML source without JavaScript.
• ☐ Add a 40 to 60 word direct answer to the 10 pages with the most impressions.
• ☐ Attach source, year and link to every number already published on the site.
• ☐ Publish pricing or the pricing model, dated, and cut the main form to six fields with one budget question.
• ☐ Set up assistant referrals as a separate analytics channel before the redesign.
• ☐ Agree the success metric with sales in writing: meetings held, not sessions.

When should you not redesign your site yet?

In four situations, a redesign is the wrong purchase and will not survive contact with your numbers.

The first is when you do not know which pages produce meetings. Without that, a redesign is a guess with a budget attached; spend two months instrumenting the funnel instead. The second is when the offer is unresolved — if positioning changes twice a year, the site gets rewritten before it is indexed, and rewriting pricing and problem pages on your current template costs a fraction of the same work on a new one.

The third is when your traffic is almost entirely paid and referral. Then the constraint is landing page conversion and cost per lead, not site architecture, and the fix is narrower and cheaper. The fourth is when your site already passes Core Web Vitals and the content is the weak point: new templates will not make a thin page quotable, the writing will. If any of these describes you, postpone the redesign and fix the specific thing that is failing.

Frequently asked questions

Is SEO dead?

No. Organic search accounted for 16.04% of total traffic in 2025 while AI referrals accounted for 0.14%, roughly a hundred to one, per a study across more than 50,000 sites by Semrush (2026). What changed is the payoff curve, not the mechanism. Optimizing for answer engines sits on top of ordinary SEO rather than replacing it: a page still has to be crawlable, indexed and ranked before an assistant can quote it. What changes is the page itself.

Should I gate my content?

Not at the top of the funnel. Content behind a form cannot be crawled, quoted or attributed, so it adds nothing to the visibility you are defending, and when an assistant summarizes your category before the buyer reaches you, being unreadable is the expensive option. Gate only proprietary material: raw benchmark datasets, working templates, tools that need an account. If you would be comfortable seeing a piece quoted with your name attached, publish it.

How fast does my site need to be?

Fast enough to pass all three Core Web Vitals: LCP at or under 2.5 seconds, INP at or under 200 milliseconds, CLS at or under 0.1, measured at the 75th percentile of real user field data over 28 days rather than in a lab test on your laptop. Only 48% of sites pass all three on mobile, according to the Web Almanac 2025 from HTTP Archive, so passing is a real differentiator. Measure on the templates that already receive traffic.

Do I need to publish pricing?

You need to publish enough for a buyer to self-qualify. Someone who cannot find a number either leaves for a competitor who published theirs or submits your form purely to extract one, which produces meetings with people who have no budget. If a rate card is impossible because every project is scoped, publish the model: scope tiers, what each includes, what is billed separately, and a typical timeline.

How long does a B2B redesign take?

Scope decides it, and the client controls most of the delay. DINOLABS publishes delivery windows of 2 to 4 weeks for an informational site and 3 to 6 weeks for e-commerce, measured from the point at which content, brand assets and access to hosting and analytics are in hand. A multi-language site, custom integrations or a CRM connection extend that. Schedules slip on content and approvals rather than on build.

Will AI assistants send me traffic?

Some, and those sessions tend to convert well because the visitor arrives further along in the decision, but the volume is small. AI referral traffic accounted for 0.14% of total traffic in 2025, up 66% year over year but from a very small base, while organic search remained at 16.04%, per a study across more than 50,000 sites by Semrush (2026). Treat it as a growing channel, not a replacement for search.