Copilot AI Referral Traffic Fell 96% From Its Peak. Did Microsoft Lose AI Search?

Copilot's 96% fall is the headline most likely to travel without its denominator.

The Previsible 2026 State of AI Discovery report followed the same 166 GA4 properties for 19 months. It measured sessions attributed to identifiable standalone LLM referrers such as ChatGPT, Gemini, Claude, Perplexity, and Copilot. It did not measure every AI-assisted journey, zero-click activity, or Google's AI search features.

That boundary changes the conclusion. The study shows where observable LLM referral traffic concentrated and which platforms gained or lost referral momentum. It cannot settle who won AI search.

What Previsible Actually Measured

All 166 properties remained in the dataset from November 2024 through May 2026. The sample covered SaaS, ecommerce, finance, legal, health, insurance, education, publishing, events and ticketing, and small businesses.

Monthly trackable LLM-referred sessions rose from 65,249 in November 2024 to 644,478 in May 2026, about 9.9 times the starting volume. ChatGPT-referred sessions rose from 47,606 to 610,910 across those endpoints.

The platform figures need equally careful wording:

• ChatGPT: 92.4% of trackable standalone-LLM referral sessions across this dataset, not 92.4% of AI search or global AI usage.
• Gemini: 18,119 referred sessions in May 2026, making it the second-largest named standalone referrer in the published May figures.
• Claude: 133 monthly referred sessions in November 2024 and 8,528 in May 2026, about 64 times the starting level. Claude passed Perplexity in March, not Gemini.
• Perplexity: down about 61% from its March 2025 referral peak of 17,507 to 6,788 in May 2026.
• Copilot: down about 96% from its August 2025 referral peak of 8,651 to 339 in May 2026.

The published May platform counts contain a small inconsistency. The five named counts total 644,684, while the report lists 644,478 sessions across all platforms for May. That 206-session difference could be a typo, revision mismatch, or an unexplained reporting detail. I would record it and avoid calculating a new May share or residual category from those figures.

Did Microsoft Lose AI Search?

No. The dataset cannot support that verdict.

A fall from 8,651 peak monthly referrals to 339 is severe for the measured channel. It says Copilot lost referral momentum across these 166 properties. It does not say Microsoft lost all AI search users, enterprise deployments, product engagement, revenue, or influence inside Windows, Microsoft 365, Azure, and its wider ecosystem.

A person can use Copilot without clicking an external result. A company can deploy Microsoft AI products without creating a GA4 referral. Copilot can influence a journey that later arrives through direct traffic, branded search, email, or another source.

Referral sessions are downstream website events, not product-usage events. I would lower Copilot's referral priority only after reviewing the site's own traffic and conversions. I would not use this study to grade Microsoft's AI business.

Claude's roughly 64-fold growth also needs its starting point. Moving from 133 to 8,528 referred sessions is meaningful, especially for technical and professional audiences, but the multiple does not erase the absolute volume. Gemini's May count is why the claim that Claude became the second-largest standalone referrer is wrong.

Google AI Search Belongs in Another Measurement Layer

Previsible excludes AI Overviews and AI Mode because Google AI search does not arrive as the same kind of standalone referrer. It also excludes AI activity that influences a user without producing a visit.

That omitted layer is large. At I/O on May 19, 2026, Google said AI Overviews had more than 2.5 billion monthly active users. This is a Google-reported product audience figure. It cannot be compared numerically with Previsible's 6.77 million cumulative referral sessions because the units, periods, products, and populations differ.

Google began rolling out the dedicated report on June 3, 2026 to a subset of website owners, initially in the UK. Access remains limited to a subset of properties. If the report is missing, the cause may be the rollout, insufficient supported-feature impressions, or eligibility. It does not automatically mean the site has zero AI visibility.

I covered the launch mechanics in my guide to Google Search Console's Generative AI performance reporting. The operational challenge now is using that report beside analytics referrers without blending two partial datasets into one fictional KPI.

What Google's Dedicated Report Shows and Withholds

The official Search Console documentation defines the dedicated metric as impressions. Google records an impression when a link to the site is shown in a supported generative AI feature on Google Search. The report currently includes AI Overviews and AI Mode, while Search Labs experiments are excluded.

A page in this report is the final linked page after redirects, with most page data assigned to Google's selected canonical URL. The safe statement is that the page's link was shown. The report does not prove Google used that page's text for a specific statement, treated it as a citation, placed it prominently, or displayed it in a particular answer context.

Google does count qualifying external-page clicks from AI Overviews and AI Mode. Its AI features documentation says traffic from these features enters the overall Search Console Performance report under the Web search type. Google's click and impression counting documentation also confirms that an external link click from an AI feature counts as a click.

The limitation is isolation. Those clicks are mixed into ordinary Web reporting, while the dedicated report exposes the AI-specific impression slice. Dividing ordinary Web clicks by dedicated AI impressions does not produce AI CTR because the numerator and denominator come from different event sets.

The dedicated report also omits the query or prompt behind an impression. Ordinary Web queries can help explain what a page ranks for, but they cannot be relabeled as hidden AI prompts. Google does not document a citation count, citation position, answer text, competitor comparison, conversion field, or separate AI Overviews and AI Mode dimension in this report.

This is an AI link-visibility report, not a complete AI attribution report.

My Ordered Audit for AI Visibility, Referrals, and Results

I use a layered audit because each system owns a different part of the journey. The datasets can be aligned by canonical page and time window for investigation, but they cannot be joined at the event level. The workflow cannot calculate AI CTR, identify every AI click, or prove an AI impression caused a conversion.

1. Lock the definitions

Create a short data dictionary before opening a dashboard. Label dedicated Google AI impressions as ai_link_impressions. Keep ordinary Search Console fields as web_impressions, web_clicks, web_ctr, and web_position. Label identifiable analytics visits standalone_llm_referral_sessions.

Do not rename Web clicks as AI clicks, a page impression as a citation, or ordinary Web queries as AI prompts. A reporting column should never acquire a stronger meaning than its source supports.

2. Export the dedicated Generative AI report

When a property has access, compare complete windows such as the latest 28 full days against the previous 28. Export linked pages, countries, devices, and dates. Keep chart and table exports because Google changes aggregation by view, so property totals and page totals can differ.

Normalize the final canonical URLs. Add page type, topic cluster, market, funnel role, owner, conversion path, and last meaningful update. That turns an impression list into a review queue.

3. Pull ordinary Web performance for the same pages and dates

Export Web impressions, clicks, CTR, average position, and page data for matching windows. Keep query exports separate and label them as Web queries.

If AI impressions rise while Web clicks fall, investigate rankings, query mix, country, device, seasonality, and SERP layout. Do not call the difference an AI zero-click rate. If both rise, you still cannot attribute the click increase to AI features. This is the central problem in AI search attribution: exposure, a shown link, a visit, and revenue remain separate events until a reliable identifier connects them.

4. Build the standalone LLM referrer view

Maintain a source group for known LLM domains and their app or subdomain variants. Report by source, landing page, page type, country, engagement, and conversion. Keep the raw referrer beside the grouped channel so naming changes remain visible.

Start with ChatGPT because Previsible's result makes it the first audit priority. Then use your own data. A developer platform may receive more valuable Claude traffic. An ecommerce site may see Gemini referrals convert differently. Previsible is a benchmark for attention, not a replacement for first-party analysis.

My earlier article on ChatGPT referral traffic and the clicky no-click future explains why more visible links can change the journey without ending zero-click behavior.

5. Connect observable landing pages to conversions

For sessions with an identifiable source, connect landing pages to leads, transactions, qualified pipeline, and revenue where consent and tracking permit. Use the source, session, campaign, and lead identifiers already present in analytics and the CRM.

Keep unattributed influence in an explicit unknown bucket. A branded search, direct visit, or sales conversation may have been influenced by an AI answer, but a plausible story is not source-level evidence. Report observable LLM referral conversions separately from conversions on pages that also received Google AI impressions.

6. Monitor a controlled prompt set

Use pages with meaningful AI impressions or sharp changes to build a small prompt set from Web queries, customer questions, site search, support themes, and sales objections. Record the exact prompt, date, country, device, feature, visible links, answer context, and competitors.

Treat each result as a dated observation. AI answers vary with location, personalization, model changes, and query expansion. A tracker can make the sample repeatable, but it does not reveal Google's full query population. My review of AI visibility prompt trackers explains that sampling limit in more detail.

What I Would Prioritize After the 92.4% Headline

First, verify that ChatGPT referrals are classified correctly and connected to landing-page outcomes. Check source naming, redirects, consent effects, and app variants before trusting the channel chart.

Next, export Google's dedicated report for every property that has access and group linked canonical pages by business purpose. High impressions on educational content may support discovery. High impressions on commercial pages may deserve closer review of conversion paths and answer context. Neither pattern proves traffic without AI-specific click data.

Keep Gemini referral traffic separate from Google AI search exposure. A visit from gemini.google.com is an observable standalone referral. An AI Overview impression belongs to Search Console's generative AI layer. A single Google AI channel destroys that distinction.

For content changes, stay close to Google's documented foundations: crawlable pages, stable canonicals, indexability, clear internal links, useful text, accurate structured data, and evidence that helps the reader. There is no special AI schema that repairs a weak page. My guide to Google's AI search documentation covers those requirements without inventing a separate technical standard.

Use Reporting Language That Preserves the Evidence

Use statements such as:

• "ChatGPT accounted for 92.4% of trackable standalone-LLM referral sessions in Previsible's 166-property dataset."
• "Google recorded 18,400 link impressions for these pages in supported generative AI Search features during the period."
• "Ordinary Web clicks for the same canonical pages rose 8%, but Search Console does not isolate which clicks came from AI Overviews or AI Mode."
• "Analytics recorded 312 sessions from identifiable standalone LLM referrers, with 11 observed lead completions."
• "Our controlled prompt set showed the page in 9 of 30 dated observations. This is a monitored sample, not a market-wide visibility score."

Avoid claims such as ChatGPT owns 92.4% of AI, Copilot usage collapsed 96%, Google AI CTR was 4.2%, or AI caused 23 sales unless another defensible source directly supports them.

Final Take

Previsible gives us a strong benchmark for the measurable referral layer. ChatGPT dominated that layer in its 166-property sample, Claude grew quickly, Gemini held second place in the published May figures, and Perplexity and Copilot lost referral momentum.

Google's dedicated report supplies different evidence: links were shown in AI Overviews or AI Mode. It closes part of the visibility gap but does not connect an impression to an AI-specific click, hidden prompt, citation context, or sale.

Use the dedicated report to find exposed pages, ordinary Web performance for aggregate Search outcomes, analytics referrers for observable standalone LLM visits, and the CRM for business results. Keep the columns separate until the source systems provide a real join.

Frequently Asked Questions

Does ChatGPT have 92.4% of the AI market?

No. ChatGPT accounted for 92.4% of trackable standalone-LLM referral sessions in Previsible's 166-property GA4 dataset. The study does not measure global AI users, search market share, prompts, citations, zero-click exposure, or Google AI search features.

Did Copilot lose 96% of its users?

No. Copilot-referred sessions in the sample fell about 96% from an August 2025 peak of 8,651 to 339 in May 2026. That does not measure Copilot's total users, adoption, enterprise usage, revenue, or every journey it influenced.

What does Google's dedicated Generative AI report show?

It shows impressions when links to a site appear in AI Overviews or AI Mode. Available dimensions are linked canonical page, country, device, and date. Access is limited to a subset of properties, and Search Labs experiments are excluded from the current documented scope.

Does Search Console count clicks from AI Overviews and AI Mode?

Yes. Qualifying external-page clicks enter ordinary Search Console Web performance. The dedicated Generative AI report does not isolate those clicks as an AI-specific metric.

Can I calculate AI CTR by dividing Web clicks by AI impressions?

No. Ordinary Web clicks include all Google Search Web activity, while the dedicated report isolates impressions from supported generative AI features. The numerator and denominator do not represent the same feature-specific event set.

Can the report show the prompt, citation, or answer behind an impression?

No. Current official documentation does not expose queries, prompts, answer text, citation context or position, competitors, or a separate AI Overviews versus AI Mode view.