Skip to main content
AI Traffic shows what AI systems and their users do with your site. Retrieval and citation evidence lead the story; referral sessions and conversions are the undercounted visible tail.

Open AI Traffic

  1. Open your project
  2. In the sidebar, open Insights › AI Traffic
  3. Stay on Overview, or select Crawls & fetches, Referrals or 404 Demand

Overview

Overview answers where demand is strong and where it leaks.

AI traffic signals

One card per stage of the journey: Where a source is not connected, the card shows no number instead of a zero, and Manage sources → opens Settings › Integrations. When findings from AI Traffic have been queued as work, a link above the cards reports how many related fixes are waiting in Actions.

Top pages

A page table with Page, Answer fetches, Referral sessions and Conversions columns, ranked by answer fetches plus referral sessions for the window.

Crawls & fetches

Crawls & fetches is the machine-traffic view. The headline and the two cards, Training crawls and Answer fetches, count verified requests: requests whose sending address DevTune confirmed against the networks the claimed operator publishes. Under each card, a muted line counts the rest as unknown (the address could not be checked) and identity mismatches (the address contradicted the claimed operator). See AI Traffic Overview for why a request can stay unknown. The reporting dates appear below the cards. If only the tracking snippet is connected, the Answer fetches card reads Server-side sensor required, because the snippet cannot see server-side answer fetches. Select Connect a sensor to add Cloudflare, edge middleware or a Vercel log drain. Unmeasured crawl or answer-fetch counts mean no requests were recorded and a server-side sensor did not cover every UTC day of the window. Referral counts are measured separately: only on days when every required AI referral host is measured. See When the project-wide count is measured. To leave a host no referral sensor can run on out of that check, exclude it under AI referral hosts in Settings › Integrations. Its referrals still count. The Answer fetches card also shows how many answer fetches came from coding agents (Claude Code, Gemini CLI, OpenCode, Qwen Code, Crush and Aider) since DevTune started counting them, with a count for each agent. The share compares them with every observed answer fetch over the same span, not the verified count, because coding agents run on their users’ machines and their fetches stay unknown. If some answer fetches in that span couldn’t be attributed to a coding agent, for example Cloudflare pulls, the card gives the attributed count on its own. Until a fetch is attributed to a coding agent, the card says none have been recorded, or, when some answer fetches couldn’t be attributed, that none were attributed. See Coding agents. Below the cards:
  • Daily crawls & fetches - one bar row per class, each bar split into verified, unknown and identity-mismatch requests
  • Platforms - Platform, Training crawls and Answer fetches, ordered by verified requests. A platform with identity mismatches carries a note with how many requests claimed it from outside its published ranges.
  • Recent answer fetches - answer-fetch records grouped by hour, with today’s answer fetches by platform

Acting agents

Acting agents are third-party agents that use your site for a buyer: booking, comparing options or filling in a form. They are neither training crawls, which collect pages for model training, nor answer fetches, which retrieve a page to answer a question. DevTune counts them separately and never adds them to the Training crawls or Answer fetches cards. When the window has acting-agent requests, an Acting agents card appears beside the other two. When a connected sensor can check request identity, it leads with verified requests and gives unknown requests and identity mismatches on the line underneath, like the other two. Without one, it shows the requests observed, with no verified split. Below that, it lists the window’s requests by platform. Only agents that send an identifying user agent are counted. Browser-based agents that don’t identify themselves aren’t, so the count is a minimum. The card is hidden when the window has no acting-agent requests, when the count is unmeasured, and when only the period before the window had requests.

Acting-agent alerts

DevTune checks acting-agent traffic once a day and alerts you in two cases:
  • A new platform. An acting agent from a platform DevTune hasn’t seen on the project before sends its first requests. Agents already sending requests when alerts started are recorded without an alert.
  • Volume. The project receives 25 or more acting-agent requests in the last 7 days. This alert is sent at most once every 7 days.
Both alerts link to Crawls & fetches, and both counts are a minimum, for the reason above. They arrive in-app and by email; turn either off under Acting agents on the Notifications tab of your user settings. To post them to Slack too, choose the project’s alert channel in Notification Channels; see Send alerts to a Slack channel.

Acting agents in the API and MCP

The Get traffic by platform endpoint and the devtune_get_traffic_platforms MCP tool return the same counts in an actingAgents block: The block follows these rules:
  • Null means unmeasured. A period with no acting-agent requests reads 0 only when a server-side sensor (Cloudflare, edge middleware or a Vercel log drain) covered the whole site on every UTC day of it. Otherwise it reads null. The tracking snippet can’t produce a zero, because acting agents fetch pages server-side. currentVerifiedRequests and currentFailedRequests are null whenever currentRequests is.
  • With a domain filter, an empty period is always null. Counts narrow to that exact host, so example.com doesn’t include www.example.com. Sensor coverage is recorded for the whole site, not per host, so a filtered period with no requests can’t be confirmed as zero.
  • The block counts page requests only. Requests for assets (images, fonts, scripts and stylesheets) and for /api/ and /_next/ paths are left out. The acting_agent rows in platforms[] count every request, assets included, so they can be higher than the block.
  • The block is omitted when it can’t be read. The rest of the response is still returned.

Referrals

Referrals leads with three metrics: Referral sessions, Conversion events and Share of visitors from AI. Sessions come from the selected analytics provider, GA4 or PostHog, and are not people. Conversion events are the selected GA4 key events or PostHog events and actions reported for AI referral traffic, not unique customers. Selected events that weren’t tied to a website visit, such as events your server sends, can’t be credited to AI, so their count is listed separately under Conversion events. Daily referral sessions charts the same window. Share of visitors from AI is the share of all sessions on your site that came from an AI answer over the same window: AI-referral sessions over all sessions, from the selected analytics provider (GA4 or PostHog). The line under the value names the provider. It is shown whenever the provider has imported every day of the window and recorded visits. Sessions that may have come from an AI answer but couldn’t be traced, because their source or site is unknown, aren’t counted as AI, and a line under the value says how many. Otherwise it reads Not measured, and says whether you need to connect GA4 or PostHog, wait for the provider to finish importing the period, or whether no visits were recorded. The change from the previous period appears only when both periods have a share. The reporting dates appear above the metrics. Recorded totals can include partial source coverage. If the analytics provider needs attention or conversions have not been selected, select Analytics in Integrations in the source notice.
  • Direct sessions per day shows GA4’s full daily direct sessions with the measured-page actual and baseline overlay. AI-correlated lift stays separate from referral counts and does not establish causation. See Direct traffic for coverage and classification rules.
  • By AI engine shows the session, conversion-event and engagement breakdown from the selected provider, ordered by conversion events and then sessions. Events per session is not the proportion of people who converted. Missing measurements remain unavailable for each engine. A link beside it opens the Measured actions tab in Results, where live actions are measured over their own before/after windows.
  • By page lists each Landing page with its Source, Sessions Δ and Conversion events, ordered by sessions. Previous, next and page-size controls move through the results; changing the reporting window starts at the first page.
Expand a landing page to see its current and previous average engagement time, engagement rate, and page value. Page value is the page’s share of AI-referred conversion events weighted by engagement rate; it is unavailable when either measure is missing. Conversion rate vs other, Bounce rate vs other and, for GA4, Engagement rate vs other compare each page’s AI-referred sessions with every other session that landed on the same page over the same window. The first figure is the rate for AI referrals; the line under it gives the difference in percentage points and the rate for the page’s other traffic. PostHog doesn’t record engaged sessions, so PostHog pages show bounce rate only. Conversion rates leave out selected events with no website visit, such as events your server sends, on both sides. A page is compared when each side has at least 30 sessions and no more than 2% of its other sessions, or of its other conversions for the conversion rate, couldn’t be traced. Each compared row shows how much of its other traffic couldn’t be traced. Otherwise the row says why: When analytics is disconnected, needs reconnecting or has no key events selected, a line above the table says so and links to the setting that fixes it. The same referral data is available from the List AI referrals endpoint and the devtune_get_ai_referrals MCP tool, which report the coverage of each count so you can tell partial counts from complete ones.

Visits per answer fetch

Visits per answer fetch shows, for each page, how many AI-referred visits arrived for every answer fetch. An answer fetch is an AI system retrieving the page while it builds an answer. It is not a citation: the answer may not have linked to or named the page. A high value means answers that read the page tend to send visitors to it; a low value means AI systems read the page but few people arrive from the answers. The table lists Page, Answer fetches, AI-referred visits, Visits per answer fetch and Referral source, up to 100 pages. Pages with complete referral data come first, highest visits per answer fetch first; pages without it follow, by answer fetches. The window is the same number of completed days as the rest of Referrals, but ends a day earlier, on the day before yesterday; the dates are shown under the description. Visits come from GA4, PostHog, the traffic snippet or edge middleware version 0.13.0 or later, and Referral source names which, or reads More than one sensor when the host’s source changed during the window. GA4 and PostHog count when the provider has imported the whole window for the page. The traffic snippet counts when it reported page views from the page’s host on every full UTC day of the window, from the day before it starts to the day after it ends. A snippet installed, removed or moved to a new key during a window doesn’t count for that window. After you install the snippet, a 7-day window first qualifies on the 9th UTC day and a 30-day window on the 32nd. Edge middleware counts when it measured the page’s host on every UTC day of the window. Answer fetches come from the crawler sensors.
  • No complete referral data for this window. No source covered the whole window for the page: the analytics provider has not finished importing it, the snippet missed a day on that host, or the analytics provider recorded AI-referred visits to the page without finishing the import, so the snippet’s count could understate them. Neither visits nor the ratio is shown.
  • N+ under AI-referred visits means a referral sensor recorded at least that many visits but did not cover every day of the window. The ratio reads No complete referral data for this window.
  • Not applicable means the page had visits but no answer fetches in the window, so there is nothing to divide by.
The project total above the table counts every page with activity, including pages beyond the first 100, except pages without complete referral data on a host excluded in Settings › Integrations: their answer fetches are left out of the total. It appears only when every other page has complete referral data. Otherwise a line says how many answer fetches are on pages without it, or that the total could not be checked for every page. The same pages and total are available from the Get visits per answer fetch endpoint and the devtune_get_visits_per_answer_fetch MCP tool.

Page performance

Open Insights › Audit and select the Pages tab to compare Google Search Console clicks with AI citations for the same pages. Select a page to see its evidence, search impressions and position, citability, and next action. Pages are recommended by opportunity; use the sort selector to order them by a metric instead.

404 Demand

The 404 Demand tab in Insights › AI Traffic is the detail view for missing-page demand. It reports a 30-day snapshot. Only a server-side crawler sensor, Cloudflare, edge middleware or a Vercel log drain, sees requests for missing pages; the site snippet does not. A project with no traffic setup at all is sent to Settings › Integrations instead, and a project with only the snippet sees no 404 demand here.

Project relevance

404 Demand ranks missing paths by assessed fit with your project, then by traffic demand. Each content finding shows whether it is relevant, adjacent, uncertain, or unrelated, with a reason. Use Sort by request volume to compare the loaded findings by their observed requests. The assessment uses your project context. Missing context, an ambiguous path, or an unavailable assessment appears as uncertain. After you change project context, prior assessments become uncertain until the next demand refresh reassesses them. Request counts and sensor coverage remain measured traffic evidence. The same relevance assessment contributes to suggestion priority alongside traffic demand and confidence. Clearly unrelated paths do not generate automatic suggestions. Relevance sets priority; it does not estimate visits or outcomes. Before drafting content, the 404 Demand triager checks existing content and whether the path fits the product.

Empty State

Until a sensor is connected, cards show no number rather than a zero. Select Manage sources → or Connect a sensor to open Settings › Integrations, connect a sensor, and wait a few minutes for data to appear.

Reporting window

Choose 7, 30 or 90 days beside the tabs. The window applies to Overview, Crawls & fetches and Referrals. A Δ column compares the selected window against the one before it.

Best Practices

Monitor Bot Crawl Patterns

Track which AI bots crawl your site most frequently. If important content pages are not being crawled, review your robots.txt to ensure AI bots are not blocked from accessing key pages.

Track Referral Growth

AI referral traffic is a lagging signal that some AI-driven visitors reached your site through attributable links. Growth in visible referrals can support the wider retrieval and visibility picture, but it does not measure total AI-driven demand.

Pair with Visibility

Use AI Traffic alongside Visibility to connect retrieval evidence with its visible downstream tail:
  • Visibility shows whether AI platforms mention and cite your content in their responses
  • AI Traffic shows what AI systems retrieve and the referral sessions and conversions that can be attributed
AI referrals are systematically undercounted because most AI-driven visitors arrive as direct traffic. Treat referrals and conversions as lagging, visible evidence rather than a complete linear funnel.

Review Top pages

The Top pages table shows which content has the most visible AI activity. Use this to:
  • Identify content with the most visible AI activity
  • Understand which types of content attract the most AI-referred visitors
  • Prioritize content updates for pages that receive heavy AI traffic

Troubleshooting

No Data Appearing

Check each sensor with Verifying your sensors, then:
  • Cloudflare pull: wait for the next hourly sync, and check the hostname is orange-cloud proxied. See If Cloudflare stops syncing.
  • Edge middleware: check the environment variables are set where you deployed, and that Snippet and middleware collection is switched on in Settings › Integrations.
  • Vercel log drain: check Show in Log Drains is on for the Vercel project.
  • Tracking snippet: check it is in the page source as copied, that Snippet and middleware collection is switched on, and that ad blockers aren’t blocking it (set up a proxy if needed).

Low Bot Traffic

  • Check your robots.txt file for rules that might block AI crawlers
  • AI bot crawling frequency varies by platform and is not under your control

Low Referral Traffic

  • Low referral traffic does not measure total AI-driven demand because most AI-driven visitors arrive as direct traffic
  • Review Insights › Visibility to see if your presence rate is high but referrals are low
  • Consider improving your content to be more actionable, encouraging users to visit your site

Next Steps