How to measure AI search in ChatGPT and Google, from the buyer's question to the sale

No AI platform shows you the whole path from a buyer's question to a sale. ChatGPT shows you some steps, and Google's AI Overviews and AI Mode show you others. For the last two steps, when the buyer visits on their own later and when they pay, neither shows you anything. Between July and October 2026, ChatGPT's live page reads fell 81%, and for a week in October its visits lost their tag and landed in Direct.

Think of an AI as a hotel concierge who recommends your restaurant. You'd want to know five things: whether the concierge read your menu, whether they recommended you over the place next door, whether the guest walked over with the concierge's card, whether the guest turned up on their own a week later, and whether they paid. ChatGPT's concierge sends some guests with a card. Google's tells you how often it mentioned you, but not which guests it sent.

In this post:

• What ChatGPT and Google's AI let you see at each of five steps, from the AI reading your page to the sale
• Why ChatGPT's bot hits fell from July to October 2026, and why its visits landed in Direct for a week in October
• What to check on your own site, the one part of the path you control

Figure 1 maps the five steps for ChatGPT and Google's AI, and marks the two steps where your own site decides what happens. Other assistants, such as Perplexity and Copilot, aren't covered.

Step 1: When AI reads your page, only ChatGPT leaves a trace

ChatGPT-User is the bot ChatGPT sends to fetch a page live while it writes an answer. On thousands of Adobe customer sites, its hits fell 81% from early July to early October 2026, while visits from ChatGPT rose 68% by late September. Hits from OAI-SearchBot, OpenAI's search crawler, rose 24% over the same months. Figure 2 puts the fall in live reads and the rise in visits on one scale.

A review of about 1.5 billion server requests found the same shift from mid-August: ChatGPT-User hits down about 80% and OAI-SearchBot hits more than doubled. The review calls this consistent with ChatGPT indexing pages in advance instead of fetching them during each conversation.

Since July 2026, ChatGPT-User hits no longer tell you how often ChatGPT used a page in its answers. Check each hit against the IP ranges OpenAI publishes, then use the hits to compare pages over the same period. The fall came in two drops, in the weeks of 20 July and 24 August 2026, so a quarter from before then isn't a fair baseline.

Google's AI leaves no trace of its own. Gemini and AI Overviews have no bot of their own, and Google's AI features get their content from the same Googlebot crawlers that fetch pages for Search.

On your site: OpenAI says sites that block OAI-SearchBot won't be shown in ChatGPT search answers, and Googlebot feeds Google's AI. Check that both reach the pages you want cited.

Step 2: Google counts your AI impressions, and prompt trackers show competitors

Google gives you one first-party number. Search Console's generative AI report counts impressions of your links in AI Overviews and AI Mode. It covers your own site only.

For ChatGPT, and for any comparison with competitors, the numbers come from prompt trackers: tools that run a fixed set of prompts on a schedule and record which brands and links each answer names.

Those numbers move a lot. Between 40% and 60% of cited domains change from month to month, and only 2.4% of cited URLs overlap across ChatGPT, Perplexity and AI Overviews, according to figures collected in a July 2026 measurement guide. Read your share of answers in a tracker as a direction over several months, and read each platform on its own.

Step 3: A click from ChatGPT carries a tag, and Google's AI report counts no clicks

ChatGPT adds a tag to the links it sends, utm_source=chatgpt.com, so your analytics can credit the visit. ChatGPT also passed a referrer, but mainly on desktop web, so 95% of mobile ChatGPT visits relied on the tag alone.

Search Console's AI report counts impressions only, so it can't tell you which visits came from an AI answer.

For a week in October 2026, ChatGPT visits landed in Direct

From 8 October, GPT-6, ChatGPT's new model, opened inline links without the tag or the referrer, so analytics filed those visits as Direct. On about 500 pages where ChatGPT drove 5% or more of visits, ChatGPT visits fell by about 31,000 a day and Direct rose by about 26,000 a day (figure 3), comparing 9 to 13 October with the week before, on sampled data. Search on the same pages fell 7%, in line with everywhere else, so the rise in Direct wasn't a wider traffic change. Most missing ChatGPT visitors were still arriving, unlabelled. In GA4 data from the growth agency Graphite, ChatGPT traffic fell by between 22% and 57%, depending on the site.

OpenAI told Cedric Huesler at Adobe on 14 October that the missing tag was a bug, and by 15 October tagged clicks were back to roughly normal levels, Relevant Audience reported. The referrer is still often missing, so count ChatGPT visits by the tag. GA4 can't add the source back to sessions it already recorded, so mark 8 to 15 October on any chart that crosses it.

On your site: across 77 sites with significant AI traffic, AI visitors converted 34% better than organic visitors on average, 49% better in B2B and 19% in B2C. Check which pages AI visitors land on and compare their conversion rate with organic. When AI visitors convert at about the same rate as organic ones, that review says it's a reason to dig in, and links weaker AI conversion to missing comparison pages, such as "X vs Y" and "best X software".

Step 4: Buyers who visit later on their own show up as brand search or Direct

Some buyers read an AI answer, don't click, and visit later on their own, like the concierge's guest who turns up a week later without a card. n8n, the workflow automation platform, found in its own data that 90% of its AI search conversions never clicked a citation link and showed as Direct. About one in three purchases in Profound's research followed a relevant AI conversation in the previous 30 days, mostly unattributed, the company said at its Zero Click event in Sydney in October 2026. It also found Google is still the most common next stop. Graphite, which reported the n8n case, and Profound both sell AI search services.

How much of AI's effect lands in brand search and Direct is unsettled. A study of 632 tests across 12 sites found no reliable link between AI visibility and later branded search or Direct traffic. In the same data, one site's Direct traffic moved with its AI share of voice, most closely about three weeks later. So I read brand search and Direct beside AI referrals as supporting evidence, not proof.

Step 5: Only asking the buyer shows AI's part in the sale

In the same n8n case study, GA4 credited AI search with about 1% of conversions, while n8n's post-conversion survey put it at about 9%.

Ask when someone enquires or in the first sales conversation, with open questions first, and only then "Did an AI assistant play any part?" A checkbox labelled "AI search" can prompt the answer you hope to hear. Sales calls carry the same signal, so tag call notes for phrases like "ChatGPT recommended you".

What I'd change in an AI search report

In my opinion, measurement is the weakest point of AI search work right now. I'd report each step and each platform on its own line, because one platform's result doesn't stand in for another's. The steps also count different things: hits, impressions, visits and people. Adding them, or dividing one by another, gives a number that means nothing.

I'd name the platform in every before-and-after claim. To separate your own work from the channel's growth, compare with a group of similar pages you leave alone. And I'd give any trend two or three quarters of data before judging it.

Ankur Lakhanpal is a search and AI discoverability strategist working across SEO, AI search (GEO and AEO), technical SEO, measurement and growth automation.