What Should a 2026 B2B AI Search Strategy Include?

Executive summary

Business buyers now ask AI tools which vendor to pick. The AI reads a few web pages and writes the answer. After that, people click less. When Google showed an AI summary, people clicked a normal search result on 8% of visits. Without one, it was 15% (Pew Research Center, March 2025).

So your website has a new job. AI has to be able to reach it, read it and trust it. That takes five things: pages the AI can get to, plain answers to buyer questions, labels (structured data) that match the page, facts with sources, and a way to count what happens.

Labels alone won’t do it. When Ahrefs tracked pages that added them, AI citations barely moved.

This guide covers the five parts in order. It also shows how to measure the results without promising what nobody can.

_{How a B2B page gets into an AI answer, and the five parts of the strategy that support it. When Google showed an AI summary, clicks on traditional results fell from 15% to 8% of visits (Pew Research Center, browsing data of 900 U.S. adults, March 2025). Structured data is one layer of five: when 1,885 pages added JSON-LD, AI citations barely moved (Ahrefs, May 2026).}

Key takeaways

• When Google shows an AI summary, clicks on traditional results fall from 15% to 8% of visits (Pew Research Center, March 2025). Software buyers moved early: 51% now start research with an AI chatbot more often than with Google (G2, 1,076 buyers, March 2026).
• A 2026 B2B AI search strategy has five parts: pages AI systems can reach, direct answers to buyer questions, structured data that matches the visible page, sourced claims kept current, and measurement against a baseline.
• Structured data helps AI systems read a page. It doesn’t win citations by itself. Ahrefs tracked 1,885 pages that added JSON-LD against 4,000 matched controls: citations changed −4.6% in Google AI Overviews, and +2.4% in AI Mode and +2.2% in ChatGPT, neither of which was significant (Ahrefs, May 2026).
• The Princeton GEO study found 30–40% more AI visibility from added structure, statistics, and citations (KDD 2024). Those were changes to the page text, tested together. The study doesn’t show what structured data does alone.
• Measure two things against a baseline. An AI citation, as we count it, is a live fetch of your page that an AI system uses to answer a question. An AI referral is a visit that arrives from an AI answer. SiftServe’s ~20% uplift figure is a hypothesis it tests in every pilot, not a promise.

This guide is for B2B marketing leads, founders and heads of growth who’ve been asked for “an AI search strategy” and want one they can defend line by line. It covers what to fix on your own website, in what order, and how to tell whether it worked. It doesn’t cover paid placement, PR or link building, and it won’t promise you a spot in any AI answer, because nobody outside the AI companies controls that.

If you want the hands-on page inspection, our guide on how to tell if your pages are ready for AI search walks through it. This one is the strategy that sits above it.

In this guide

• The 2026 strategy at a glance
• What AI search visibility means
• What the evidence does and does not show
• Make important B2B pages accessible
• Answer buyer questions on the page
• Use structured data to clarify not replace content
• Support claims and maintain an approved source
• Measure and revise
• Frequently asked questions
• Where to start

The 2026 strategy at a glance

A B2B AI search strategy is a plan for making the pages buyers rely on easy for AI answer engines to reach, read, trust and cite, and for measuring whether they do.

Start with accessible, accurate pages that answer the questions buyers actually ask. Add structured data that describes those pages honestly. Back every claim with a source. Then measure AI citations and AI referrals against a baseline you recorded before you changed anything.

We call the order reach, answer, label, prove, measure. It matters more than any single step, because each layer depends on the one before it. Structured data on a page an AI crawler can’t reach does nothing. A sourced claim buried in a PDF does very little.

Part	What it does	What it can’t do
1. Reach: accessible pages	Lets AI crawlers and AI agents fetch the pages buyers need, with the content in the delivered HTML	Make an AI system choose your page
2. Answer: direct answers	Gives an answer engine a sentence it can lift: what you sell, for whom, at what price, against which criteria	Replace evidence or reputation
3. Label: structured data	Labels what the visible page already says (organization, product, article, FAQ)	Win citations by itself
4. Prove: sourced, current claims	Makes your facts attributable and safe to quote	Fix a page nobody can read
5. Measure: citations and referrals	Shows whether citations, referrals and AI answers changed against a baseline	Prove cause from a single before-and-after

Structured data alone doesn’t secure AI visibility. It’s one layer of five, and the table puts it third for a reason.

What AI search visibility means

Four terms do most of the work in this guide. They get used loosely, so here’s how we use them.

AI answer. A response written by a language model (ChatGPT, Google’s AI Overviews and AI Mode, Perplexity, Gemini, Claude, Copilot) after it reads web sources, in place of a list of links. The buyer gets a recommendation or a comparison without opening your site.

AI citation. We count a citation whenever an AI system fetches your page live to answer someone’s question and uses what it read in the answer. The question can be about your brand, about the buyer’s problem, or about anything your page helps answer. A visible link in the answer isn’t required. If the page was fetched to build the answer, it counts. You can see these fetches in your server or edge logs, because the assistants use on-demand agents for them. OpenAI’s ChatGPT-User, for example, visits a page “when users ask ChatGPT or a CustomGPT a question” (OpenAI crawler docs).

AI referral. A human visit that arrives at your site from an AI answer: someone clicked a source link in ChatGPT, Perplexity or another assistant. Analytics shows these by referrer. Citations tell you the AI read you. Referrals tell you a buyer followed.

Structured data. Machine-readable labels in the page code, usually Schema.org vocabulary written as JSON-LD, that say what the page is about: this is an organization, this is a product with this price, this is an article by this author on this date. People never see it. Search engines and some AI systems read it.

Visibility, then, is the combination: how often AI systems fetch and use your pages, how accurately their answers describe you, and how many buyers click through. None of the three is guaranteed by anything on your page. All three can be measured.

What the evidence does and does not show

There are four findings worth knowing here. Each one supports a narrower claim than it usually gets credit for.

Clicks are falling when AI answers appear. Pew Research Center analyzed the browsing data of 900 U.S. adults in March 2025. When Google showed an AI summary, users clicked a traditional result on 8% of visits. Without one, the figure was 15% (Pew Research Center, July 2025). This is consumer behaviour on Google, and it says nothing about which pages got cited. It does tell you that being the source of the answer matters more than it used to.

B2B software buyers are already there. In G2’s survey of 1,076 B2B decision makers, 51% said they start software research with an AI chatbot more often than with Google, and 69% said a chatbot surfaced information that led them to a different vendor than they expected (G2, March 2026). G2 sells software reviews, and the sample is software buyers, so read it as a strong signal for SaaS rather than a law for all of B2B.

Content changes lift AI visibility, together. The Princeton GEO study (Aggarwal et al., KDD 2024) found 30–40% more AI visibility from added structure, statistics, and citations (arXiv). The methods it tested were edits to the text of a page: adding sources, quotations and statistics, and making the writing easier to follow. It didn’t test Schema.org markup. So the finding supports “write citable pages”. It doesn’t support “add schema and get 30–40% more visibility”, which is how it often gets quoted.

Adding schema alone barely moved citations. This is the newest piece. Ahrefs started from 6 million URLs and found that AI-cited pages were almost three times more likely to carry JSON-LD than pages that weren’t cited. Then they tested whether that was cause or coincidence. They tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 and compared them with 4,000 matched pages that didn’t (Ahrefs, May 2026). Thirty days later, the citation changes were:

Platform	Change in citations vs matched controls	Statistically significant?
Google AI Overviews	−4.6%	Yes (a small decline Ahrefs couldn’t explain)
Google AI Mode	+2.4%	No
ChatGPT	+2.2%	No

The correlation was real. Ahrefs reads it as a sign of the kind of site that bothers with schema, which tends to be better maintained and more technical. Ahrefs lists its own limits. Every page in the study was already heavily cited (100+ AI Overview citations each in February 2025), the window was 30 days, and all schema types were pooled. So it can’t tell you whether schema helps a page AI systems have never picked up. It does tell you that adding schema on its own is unlikely to make citations follow.

_{Schema alone vs content changes together. Ahrefs (May 2026) tracked 1,885 pages that added JSON-LD against 4,000 matched controls: −4.6% in AI Overviews, +2.4% in AI Mode and +2.2% in ChatGPT, the last two not significant, even though AI-cited pages were almost three times more likely to carry JSON-LD across 6 million URLs. The Princeton GEO study (KDD 2024) found 30–40% more AI visibility from added structure, statistics, and citations. Different studies and different measures, shown side by side, not on one scale.}

Together, the four findings set the order of the strategy. Buyers are getting answers instead of clicks. What earns a place in those answers is content an engine can read and trust. Markup helps the engine read, and the visible content still has to make the case.

Make important B2B pages accessible

Most B2B sites have maybe ten pages that do the selling: the product pages, pricing, integrations, security and compliance, the main use cases, a comparison or two. Start there. Check two things for each one.

First, can the right bots get in? Look at robots.txt, your CDN or firewall rules, and any bot-protection product in front of the site. The AI companies run several agents with different jobs, and they’re easy to mix up. OpenAI alone has three (OpenAI crawler docs):

• GPTBot crawls for model training. Blocking it says “don’t train on this”.
• OAI-SearchBot surfaces sites in ChatGPT search. Block it and you’re out of ChatGPT’s search answers.
• ChatGPT-User fetches a page live when a person asks a question. OpenAI notes robots.txt rules may not apply to it, since a user started the request.

It’s easy to block GPTBot for a good reason and catch every OpenAI agent with the same rule. Decide each one on purpose.

Second, does the page carry its content when a bot gets it? No major AI crawler executes JavaScript: across 500M+ GPTBot fetches, Vercel and MERJ found zero JavaScript execution (Vercel × MERJ, December 2024). So whatever appears only after scripts run, the crawler never sees. The problem is common. 57% of top sites show AI crawlers a nearly empty page, 403 of the 711 readable sites in the top 1,000 (ModPageSpeed, May 2026).

The quick test: fetch the page without a browser (or open view-source) and search for your one most important sentence. If it isn’t in there, nothing further down this list helps that page yet. Our readiness guide has the full seven-check inspection.

Answer buyer questions on the page

B2B buyers ask an AI assistant the questions they’d ask a sales rep before they’ll talk to one. Which tools do this for a 200-person company? Does it integrate with Salesforce? Is it SOC 2 compliant? Roughly what does it cost? What’s the difference between you and the other one?

If your pages don’t answer those questions in plain sentences, the AI answers them from somewhere else. Often that’s a competitor’s page, a review site or a two-year-old forum thread.

What works:

• Lead with the direct answer. The first two sentences of a product page should say what it is, who it’s for and what problem it solves, in words a buyer would use. “Workflow automation for mid-market finance teams” can be quoted. “Reimagine how work flows” can’t.
• Use the buyer’s terms, then yours. If the market says “accounts payable automation” and you say “spend orchestration”, say both, and define yours the first time it appears.
• Write down the evaluation criteria. Pricing model (even a range or a starting point), deployment time, integrations, security certifications, support terms, the company size you fit. These are the facts AI answers compare on. If they live in a sales deck, they don’t exist for an answer engine.
• Say who it isn’t for. It reads as confident, and it gives the AI a reason to recommend you for the right buyers instead of vaguely for everyone.
• Add an FAQ in the buyer’s words, with each answer complete on its own. That’s the format an engine can lift whole.
• Keep facts as text. A pricing table saved as an image or a compliance badge with no text beside it is invisible to a crawler.

None of this is new to anyone who’s written a good sales one-pager. The difference is that the one-pager now has to live on the public web, in HTML, on the page the question lands on.

Use structured data to clarify not replace content

Structured data is worth doing, as long as nobody sells it to you as the strategy.

Google’s own guidance is blunt. To appear in AI Overviews or AI Mode, “there are no additional technical requirements” beyond being indexed and eligible for a snippet, and “there’s also no special schema.org structured data that you need to add” (Google Search Central). The same page lists “making sure your structured data matches the visible text on the page” as a best practice. Work by that rule. Markup describes what a reader can already see on the page, and it never adds facts the page doesn’t state.

Schema type	What it can clarify	Where it goes wrong
Organization	Your legal name, logo, official profiles, contact, so mentions resolve to one company	A name or URL that differs from the one on the page
Product / SoftwareApplication	What the product is, its category, an offer or price	A price in the markup that isn’t on the page, or is out of date
Article / BlogPosting	Author, publish and update dates	A “last updated” date bumped with no real change
FAQPage	Which questions the page answers	Questions in the markup that aren’t visible on the page
BreadcrumbList	Where the page sits in your site	Paths that don’t match your navigation

Two practical points. Most CMS platforms and SEO plugins generate the basics, so check what’s already there before anyone writes JSON-LD by hand (paste a URL into Google’s Rich Results Test or the Schema.org validator). And put the markup in the HTML the server sends. If a tag manager injects it with JavaScript, the AI crawlers that don’t run JavaScript never see it.

Support claims and maintain an approved source

An AI answer engine is in the business of repeating facts. Give it facts it can repeat safely.

That means three habits. Attribute every number: “clicks on traditional results fall from 15% to 8% of visits when Google shows an AI summary (Pew Research Center)” can be quoted, and “AI is killing clicks” can’t. Date anything that changes, like prices, customer counts, certifications and benchmarks. And keep one approved version of each fact, so your homepage, pricing page, docs and LinkedIn don’t contradict each other. When two of your own pages disagree, a model has no way to pick, and it may quote the wrong one.

This is the part of the strategy SiftServe is built around. SiftServe is web infrastructure for AI agents. It creates an approved, AI-readable version of your existing website (the same facts, restructured as semantic HTML with clear headings, FAQ pairs and structured data that matches the text) and serves it to AI and bot traffic at the edge. Human visitors keep seeing your site exactly as it is. You review and approve every page before it ships, and each claim on the sifted page traces back to your original. So the approved source is the copy AI agents receive, with no separate document for someone to keep in sync.

We’ve run this on our own site since July 2026. On the published version of our homepage, the CORE-EEAT audit score moved from 65 on the original to 76 on the sifted page (methodology). Running it also taught us the lesson that matters most here. In August 2026 a configuration regression had our edge serving bots a stale copy of the homepage while humans saw the current one. Nothing looked wrong in a browser. We caught it because we routinely fetch our own pages the way GPTBot does and compare. Whatever approach you take, fetch your pages the way a crawler does and read what comes back.

SiftServe doesn’t promise visibility. Our working hypothesis is that sifted pages earn about 20% more citations and AI referrals than their unsifted baseline. We stated that before the data, and every pilot tests it. If you want a starting read first, the free AI visibility checker scores a domain from public signals in a few minutes, with no sign-up.

Measure and revise

You can’t manage AI visibility from a gut feel or a screenshot of one ChatGPT answer. Answers vary between runs and between weeks. So measure the way you’d measure any channel: baseline first, then change, then compare.

What to measure	How	What it tells you
AI citations	Count live fetches of your pages by on-demand AI agents (ChatGPT-User and its equivalents) in server or edge logs	Whether AI systems are reading your pages to answer questions
AI referrals	Segment visits from AI assistants by referrer in analytics	Whether buyers follow the answer to your site
Answers to grounding prompts	Run a fixed set of buyer questions in the main assistants on a schedule and record what each answer says	Whether you appear, and whether the description is right
Page readability	Score each page before and after a change on the same rubric	Whether the page itself improved

A few rules keep the numbers honest.

Record the baseline before you touch anything. Pick the pages you’re changing, and pull citations, referrals and the grounding-prompt answers for a fixed window first. Then compare windows of the same length afterward. Better still, leave some comparable pages unchanged and compare the two groups. That’s how our pilots work: eight weeks, citation increases measured against the unsifted baseline, and the answers to each grounding prompt tracked over time.

Write grounding prompts from real buyer questions. Pull them from sales calls, support tickets and the questions on your FAQ. Include ones that name your brand, ones that only describe the problem, and comparison questions. Keep the wording fixed so the runs are comparable.

Watch what the answers say, not only whether you’re in them. An answer that cites you but gets your pricing model wrong is a content problem you can fix this week.

Revise from what you see. If a grounding prompt keeps pulling a competitor’s comparison page, you probably need a comparison page. If citations rise but referrals don’t, the answer may be complete enough that nobody needs to click, and your page’s job then is to make sure that answer is accurate. Re-measure on the same schedule, with the same prompts.

And keep “readable” and “cited” in separate columns. The first is something you fixed. The second is something you’re measuring, and it’s affected by a lot that isn’t your page.

Frequently asked questions

Is structured data enough for AI search visibility?

No. Structured data labels what a page says, which helps machines read it, but it doesn’t get a page cited on its own. Ahrefs tracked 1,885 pages that added JSON-LD against 4,000 matched controls and found citations moved −4.6% in Google AI Overviews and +2.4% and +2.2% in AI Mode and ChatGPT, the last two not significant (May 2026). Google says no special schema is needed for its AI features. Do it, keep it accurate, and put your effort into the visible content.

Which part of the strategy should a B2B team tackle first?

Access, on the five to ten pages buyers’ questions land on: product, pricing, integrations, security and your main use cases. Confirm the AI crawlers you want aren’t blocked and that the key content is in the HTML the server sends. A page an AI system can’t fetch can’t be cited, so this unblocks everything else. Then write direct answers to buyer questions on those same pages. Structured data and measurement come after.

How do you count an AI citation if the answer doesn’t show a link?

We count the live fetch. When someone asks an assistant a question and it fetches your page to build the answer, that’s a citation, whether or not a link appears. These fetches show up in server or edge logs under the assistants’ on-demand agents, such as OpenAI’s ChatGPT-User. A fetch tells you the page was read for an answer. The grounding prompts tell you what the answer did with it.

Should we block GPTBot?

Blocking GPTBot is a training decision, and it doesn’t take you out of ChatGPT search. GPTBot crawls for model training. OAI-SearchBot is what puts sites in ChatGPT search, and ChatGPT-User fetches pages live when a person asks a question (OpenAI crawler docs). You can block training and still allow search. Check that a broad rule in robots.txt or your firewall isn’t blocking all three.

Do we need an llms.txt file?

It won’t hurt, but don’t count it as a strategy. Ahrefs found 97% of llms.txt files received zero requests in May 2026, across 137,000 domains (Ahrefs, June 2026). Our comparison of SEO, GEO, llms.txt, schema and AI-readable websites covers where each one fits.

Where to start

So what should a 2026 B2B AI search strategy include? Five parts, in this order: pages AI systems can reach, plain answers to the questions buyers ask, structured data that matches the page, claims with sources that you keep current, and measurement against a baseline. Structured data is one of the five.

This week, list the ten pages that do your selling and fetch each one the way a crawler does. Write down twenty questions your buyers ask and run them in ChatGPT, Perplexity and Google’s AI Mode. Save the answers with the date. That’s your baseline, and it costs nothing.

Then fix in order and re-measure on the same prompts. If the pages that matter are hard to change in place, our guide to making a website AI-readable without rebuilding it covers the options, and the eight-week pilot measures citations and AI referrals for sifted pages against the ones you leave alone.