Are ChatGPT, Claude or Perplexity sending you visits? How to measure them
In 60 seconds
- GA4 can identify sessions that retain an assistant referral and show where they landed and what they did.
- Search Console measures Google: its standard report covers organic search, while its generative AI report lets you review AI Overviews and AI Mode impressions.
- Logs record requests to your website, including bots. An automated fetch does not equal a sales visit or a recommendation.
- Gaps remain: answers without links, mentions without clicks and visits whose source is lost. Record these limits before comparing channels.
Official sources and feature availability checked on October 30, 2026. Examples are synthetic and do not represent client data or results.
For a B2B business owner, the useful question is which pages attract inquiries the team can handle. Answering it requires keeping an appearance, a session and a quote request separate.
What each source answers
| Source | What you can observe | Decision it supports |
|---|---|---|
| GA4 | Attributed sessions, landing pages and measured events | Improve the page receiving visits and the path to contact |
| Search Console | Appearances and clicks in Google; visibility in its AI features | Review content that appears for relevant searches |
| Server or CDN logs | Requests, URL, HTTP status and declared agent, when recorded | Diagnose access errors or pages receiving crawls |
Keep three separate sheets or views. Adding impressions, sessions and HTTP requests produces a number with no commercial meaning.
GA4: find the source and landing pages
Open Reports > Acquisition > Traffic acquisition and select Session source / medium. This dimension describes the session; avoid substituting first user source when asking where this month’s visits came from. Google’s traffic acquisition documentation explains the scope of these dimensions.
Look for assistant sources. Synthetic examples of values worth checking include chatgpt.com / referral, claude.ai / referral and perplexity.ai / referral. The app, browser and link parameters can change what arrives. An empty table does not prove there were no AI visits.
Create a custom AI channel
With Editor permissions or above, go to Admin > Data display > Channel groups. Copy a group and add an AI channel. Place it above Referral: GA4 uses the first matching channel. Google provides an official AI assistant example.
As an initial rule of our own, to review with your Analytics administrator, combine with AND:
- Medium exactly matches
referral. - Source matches this regular expression:
^([a-z0-9-]+\.)*(chatgpt\.com|chat\.openai\.com|claude\.ai|perplexity\.ai)$This list deliberately covers these assistants, including ChatGPT’s former domain. Test it against your actual values. If a source arrives through UTMs with a different name, document an additional rule. Keep traffic tagged as advertising in paid channels: this example only groups referrals. Select your custom group when reading reports; you have not rewritten the default group or recovered missing source data.
Cross the referral with the landing page
In Explore > Free form, import Landing page + query string, Session source / medium, Session source and Session medium, plus Sessions and Key events. Put the landing page in rows and source/medium in columns; filter source using the expression above and medium to referral. The free-form exploration guide documents tables and filters. The landing page report explains that a landing page is the session’s first page, rather than every page viewed.
Agree on the event that represents a successfully submitted inquiry and mark it as a key event. A WhatsApp click only establishes that click; check whether an inquiry arrived and whether it was qualified with sales or the CRM. Keep personal data out of URLs and events.
Synthetic example, without figures: referrals reach a technical product page, but nobody requests a quote. First test the button on mobile and check whether the page explains coverage, applications and information needed for a quote. The source alone does not explain abandonment.
Why GA4 can undercount
Consent mode changes collection: basic implementations block tags without consent; advanced implementations can send cookieless signals. Blockers or broken tagging can also leave visits out. Behavioral modeling has data prerequisites; a small business should not assume it will qualify or treat estimates as observed visits.
Record banner and tagging changes alongside the monthly report. A coinciding drop calls for checking measurement before blaming the channel.
Search Console: start with Google and distinguish the AI view
If you have only used Analytics, request access to the correct Search Console property and verify that it covers your domain. Reading it does not require changing website platforms.
In the Search performance report, an impression records an appearance under Google’s counting rules; a click is a departure to your website. CTR relates clicks to impressions. Average position summarizes the position of your site’s topmost result; it is not a fixed ranking. The queries and pages tabs help identify topics and URLs to review.
A dedicated generative AI view exists on the review date. Google announced it in June 2026 and confirmed worldwide rollout by August 31. Its current help page documents AI Overviews and AI Mode impressions by page, country, device and date. If the report is missing, there may not be enough impressions.
This view currently does not provide clicks, CTR or queries. These features’ data also contributes to overall Search performance: do not add the two reports together. Nor should you attribute all clicks on a page to AI because it appears in both. Search Console does not measure referrals from ChatGPT, Claude or Perplexity.
Logs: who requests your pages and why
Ask your hosting provider for a monthly access-log summary. If a CDN serves cached pages, check its logs: the origin server may not see those requests. Request date, path, HTTP status and user-agent; agree on access and retention without circulating complete logs containing IP addresses or personal query parameters.
These identifiers are documented by their providers, checked on October 30, 2026. They are tokens to find within the full user-agent; versions can change.
| Provider and official source | Training | Search | User-initiated request |
|---|---|---|---|
| OpenAI | GPTBot | OAI-SearchBot | ChatGPT-User |
| Anthropic | ClaudeBot | Claude-SearchBot | Claude-User |
| Perplexity | No training bot in this official list | PerplexityBot | Perplexity-User |
Training means collecting content that could be used to develop models. Search means crawling to help find sources. User agents retrieve pages following people’s actions; those requests still do not establish a human visit to your website. Perplexity states that neither listed agent collects content for foundation-model training. OpenAI also documents OAI-AdsBot for reviewing advertising destinations; keep it separate from search fetching.
A user-agent can be spoofed. Compare the originating IP with current ranges published by the provider where available, following its verification method. If all you have is the agent text, label the identity as unverified. Do not confuse requests with unique pages or turn fetches into citations: logs do not show the answer the person received.
What these sources still cannot tell you
You will not get a census of mentions without clicks across all assistants, or of answers without links. A page can be cited from an index or an earlier copy without a fresh request appearing in your logs.
When a referral is lost, some traffic can appear as direct or under another attribution depending on the journey. Google defines direct as traffic without a clear referral source. Do not relabel its increases as AI based on intuition. You can ask how someone found you in a form, but keep that self-reported answer separate from technical attribution.
Make content readable without promising citations
Check that product names, specifications and links arrive in HTML without waiting for an animation, scroll or click. Google executes JavaScript, but not every bot does. A heavy animation does not automatically exclude a page from the index; inspect the received HTML, rendered content and mobile performance before deciding.
Use structured data that describes visible content and review whether robots.txt or security systems prevent intended access. Google’s guide to generative AI features retains SEO foundations and does not guarantee indexing or appearance.
llms.txt offers context and links to agent-readable content. Its reference website presents it as a proposal, currently v2, revised in August 2026. Publishing it does not guarantee an assistant will read it, cite you or send traffic; it does not replace access permissions or measurement.
A static site can simplify delivering complete HTML, and a CMS or ERP website builder can deliver it too. For a migration decision, compare operations and maintenance in Odoo, WordPress or a static site. Changing technology just to add llms.txt does not solve attribution gaps.
A monthly routine that ends with a decision
| What to review | What to record | Practical decision |
|---|---|---|
| GA4: AI channel, source/medium and landing page | Sessions and verified inquiries; tagging changes | Fix a page that attracts interest but fails at contact |
| Search Console: queries and pages | Relevant topics with impressions and clicks | Complete a product page that leaves a buyer’s question unanswered |
| Search Console: generative AI view | URLs with impressions and their trend | Review the accuracy of content appearing in Google |
| Logs: verified bots, paths and status codes | Repeated errors and CDN/origin coverage | Request technical diagnosis of failing access |
| Sales team or CRM | Qualified inquiries and known, self-reported or unknown source | Prioritize improvements based on real opportunities |
Compare equivalent periods, retain the channel definition and assign one improvement for the next month. With small volumes, examine cases and journeys before interpreting percentages.
For tool choices, return to the comparisons hub. For paid acquisition, read the ChatGPT advertising guide for Latin America: appearing as a source in an answer and paying for an ad are different routes that need separate measurement.
To review how an inquiry travels from your website to sales, book a conversation with Kiia. We can start with one page and a test journey.
Frequently asked questions
Can I count how often ChatGPT recommends my business without a click?
GA4 and server logs do not provide that total. A bot request does not prove a recommendation either. Search Console reports impressions in Google's AI features, but does not measure answers from ChatGPT, Claude or Perplexity.
Does Search Console now provide a dedicated view of AI Overviews and AI Mode visibility?
On October 30, 2026, Google Search has a dedicated generative AI performance report. It shows impressions by page, country, device and date. The reviewed documentation does not provide clicks, CTR or queries in that view; it is not a report of visits from other assistants.
Do I need a static website or llms.txt to measure AI traffic?
No. Measurement depends on tagging, available source information and access to logs. A static site can make delivering complete HTML easier, but other platforms can do that too. llms.txt remains a proposal and does not guarantee visits or citations.
From insight to action
Want to turn this into an agent that works for your team?
Tell us which process you want to improve. In a free call, we will identify the first workflow worth building.