AI Search Is Changing B2B Research. Here's What We Can Measure.

AI-generated search is a real discovery channel. The honest headline is not that it has consumed one universal percentage of B2B research; it is that Google, Microsoft, and OpenAI now give site owners concrete ways to participate in and measure AI-assisted discovery.

Headline adoption percentages, universal conversion multipliers, and claims that most B2B research has already moved can sound decisive while hiding major differences in audience, survey design, attribution, industry, and time period. Those figures should not be treated as universal benchmarks without the underlying primary research.

Your firm’s own data is more useful.

What has actually changed?

Search products increasingly generate a synthesized response and provide links to supporting pages. Google says its AI Overviews and AI Mode use core Search systems to retrieve current pages from its index. It also says the same SEO fundamentals remain relevant. ¹

Bing now exposes AI citation activity in Webmaster Tools. OpenAI documents how publishers can make content available to ChatGPT search and how ChatGPT referral URLs are tagged. ^{2 3}

Those are first-party product facts. They establish that AI search deserves measurement. They do not prove that every professional services buyer has abandoned Google, that every AI visitor is pre-qualified, or that a particular engine will cite a page on schedule.

What can a B2B firm measure today?

Google visibility

Use Search Console to monitor indexed pages, queries, impressions, clicks, and landing pages. Google announced dedicated generative-AI performance views for Search and Discover in 2026, initially rolling them out to a subset of sites. Use those views if they are available in your property; otherwise, keep the overall Search performance report as the baseline. ⁴

Bing and Copilot citations

Bing’s AI Performance public preview reports total citations, average cited pages, sampled grounding queries, page-level activity, and trends across supported Microsoft AI experiences. Bing cautions that a citation count does not indicate placement, authority, or the role a page played in an answer. ²

That caveat is important. A citation is a visibility event, not a lead.

ChatGPT referrals

OpenAI says ChatGPT search referral URLs include utm_source=chatgpt.com. Track that source in analytics, then evaluate landing-page engagement, form starts, qualified inquiries, and closed work using the same standards applied to other channels. ³

Do not assume all AI-assisted journeys will be attributable. A buyer may read an answer, return later through branded search, or contact the firm offline.

Repeated answer observations

Maintain a small set of questions that reflect real client research:

• problem-aware questions
• service-comparison questions
• local or jurisdictional questions
• questions about process, timing, and cost factors
• branded questions about the firm and its professionals

Record the exact prompt, surface, date, location or account context, cited sources, and whether the answer was accurate. Repeat the test. Generated responses can vary, so a single result is an observation, not a stable rank.

Which metrics belong on the scorecard?

Layer	Metric	What it tells you
Access	Crawl status, indexed URLs, sitemap health	Whether systems can discover the content
Search visibility	Impressions, queries, clicks, landing pages	Whether the site appears and earns visits in search
AI visibility	Cited pages, citation counts, sampled grounding queries	Whether supported AI surfaces reference the site
Accuracy	Correct name, services, people, location, qualifications	Whether the generated representation is trustworthy
Engagement	Qualified sessions, form starts, calls, downloads	Whether visits show useful intent
Business outcome	Qualified leads, opportunities, retained work	Whether the channel contributes to the pipeline

Avoid collapsing these into one proprietary “GEO score.” A page can be crawlable but not cited, cited but not clicked, or clicked without producing a qualified inquiry. Each stage calls for a different fix.

What content is most defensible?

Publish content based on knowledge the firm genuinely has: original analysis, named methods, clear examples, current primary sources, and explanations reviewed by qualified professionals. Google’s people-first guidance emphasizes first-hand expertise, a clear audience, and content that helps the reader accomplish a goal. ⁵

The original GEO research supports a narrower point. In its experimental benchmark, adding relevant citations, quotations, and statistics could improve source visibility, with results varying by domain and method. It does not demonstrate a permanent cross-platform ranking boost or a guaranteed commercial return. ⁶

What should a firm do this quarter?

1. Establish access. Check indexing, sitemaps, canonicals, robots rules, and OAI-SearchBot access.
2. Create a baseline. Export Search Console and Bing data, configure analytics channel rules, and save repeated prompt observations.
3. Improve the highest-value pages. Clarify the service, audience, location, author, evidence, and next step.
4. Publish original expertise. Replace commodity summaries with professional analysis that a buyer could not get from a generic source.
5. Review monthly. Compare citations and traffic with qualified actions, then update pages where the evidence is weak or the generated description is wrong.

AI search is changing research behavior, but the size and commercial value of that change will differ by firm. Measure your market, improve what users actually need, and let observed outcomes guide investment.

Footnotes

1. Google Search Central, Optimizing your website for generative AI features on Google Search. ↩

2. Microsoft Bing, Introducing AI Performance in Bing Webmaster Tools. ↩ ↩²

3. OpenAI, Publishers and Developers FAQ. ↩ ↩²

4. Google Search Central, Introducing Search Generative AI performance reports in Search Console. ↩

5. Google Search Central, Creating helpful, reliable, people-first content. ↩

6. Aggarwal et al., GEO: Generative Engine Optimization. ↩