Skip to main content
Results shows what happened after actions went live and after detected content changes. It uses page-level before-and-after measurements and reports what moved, not proof that an action caused the change.

Where results appear

  • Results in the project sidebar lists every measured action in the project, and content changes that no action claims.
  • Each live action’s page shows its own result, so you can read it where you did the work.

Open Results

  1. Open a project
  2. Select Results in the project sidebar
  3. Choose Measured actions or Unclaimed changes
Use the 7 days, 30 days, or 90 days control to choose which measured actions or changes appear. Row deltas always compare the 14 days from the date the action went live with the 14 days before it. Claimed crawl changes retain their recorded change date; the reporting-window control does not change that comparison.

Measured actions

The Measured actions tab is the default view. Its scoreboard summarizes:
  • Conclusive measurements: actions whose complete comparison has usable evidence
  • Improved: conclusive actions where a metric rose and none fell
  • Improvement rate: improved actions divided by conclusive measurements, unavailable when there are none
  • Still measuring: actions whose comparison window is still open
  • Unmeasured: actions without a usable comparison
DevTune compares the 14 days after the change with the 14 days before. A metric counts as moving when its count changes further than chance would explain:
  • Citations count the probe answers that cite the page. The comparison accounts for how many probe answers each window contained. Method v2 uses only prompts with live, usable answers on the same platform in both windows. Adding or removing prompts does not change the v2 comparison cohort.
  • Answer fetches count the requests AI systems made for the page. Method v2 limits exposure to sensor-covered days.
  • A move counts when a shift that large, up or down, would happen by chance less than 1 in 10 times, the higher window holds at least 3, spread over at least 3 days, and the rate moves by at least 20%. A one-day burst, or a small wobble on a busy page, doesn’t count. Method v2 allows for variation between prompts as well as between days. A citation change needs at least three prompt and platform pairs with answers in each window. Large swings can keep a change inconclusive.
Under method v2, a metric needs full coverage of both windows to conclude that it rose or fell. Partial coverage reads No clear change; when neither metric has comparable evidence, the result is Unmeasured. Newly closed windows use method v2. Previously closed windows keep their v1 result. The ledger shows Method changed when its rows mix methods. Method v2 does not apply corrections that arrive after the window closes. The API and MCP ledger return items[].method as v1 or v2 for closed Results windows with evidence. Open windows and Agent Picks rows return null. Agent Picks results carry their own method in agentPicks.method. The same rule applies in both directions. Hover over a result to see what moved for that action, and how often chance would move it that far. Citation comparisons require probe evidence in both windows. The citation-lift qualifier sums citation improvements only. For the same page, overlapping before/after windows contribute only the earliest supported comparison in each overlapping group. Each action keeps its individual result. A qualifying rise does not prove that the action caused it, and a qualifying fall does not prove the action harmed the page. Each row shows the action, result, citations per day, presence change, answer fetches when available, and measurement date. Select a row to inspect its evidence. Use the result filter and page-size control to explore the ledger. Previous and Next preserve the selected range and filter across reloads.

Result and confidence

  • Improved: at least one metric rose and none fell
  • Mixed: one metric rose and another fell
  • No clear change: the closed comparison has usable evidence, but no metric moved in either direction
  • Declined: at least one metric fell and none rose
  • Still measuring: the comparison is pending
  • Unmeasured: no usable comparison exists; missing evidence never counts as a decline
A citation improvement can be reported before the answer-fetch window closes. If answer fetches then fall, the result becomes Mixed. Confidence describes sensor coverage, not causality:
  • observed — the sensor covered the full measurement window
  • directional — the sensor covered part of the window
  • still measuring — the measurement window or its data is not ready
  • unavailable — no comparable sensor evidence exists
Cells use states such as measuring · day N of 14 and window open instead of treating incomplete measurement as zero.

Measured action detail

Select a ledger row to review:
  • The lifecycle from Action completed through Result available
  • Citation and presence changes
  • Answer-fetch changes and their sensor basis
  • AI referrals as unavailable when no comparable GA4 or PostHog measurement exists
  • A daily page-signals chart with the detected-change marker
  • The 14-day comparison for actions marked live; recorded content changes also offer 7-, 30-, and 90-day windows
  • Content hashes, crawl evidence, and a link to page history
When other actions targeting the same page were completed during the comparison’s before or after period, the measured-actions list shows a count. The measurement detail lists their titles, completion dates, and links to their results for each page. When actions overlap, the results reflect the page’s overall performance and cannot be attributed to a single action. If the overlap check is unavailable, the measurement remains available and shows that the check could not be completed. Open windows remain labeled window open until enough time has passed. Page charts and longer impact windows receive citation and presence updates after new probe results arrive, through day 90 after the measurement starts. Results remain available afterward. Method v2 fixes the classification at window close. Method v1 retains its existing updates from late citation and presence evidence for days whose citation and mention detail is still kept, which is 180 days; see How long history is kept.

404 requests after a redirect

For a missing-page action shipped with a redirect, Results shows 404 requests to the missing page, with the requested path and before/after counts from the same answer-fetch comparison windows.
  • Stopped means there were 404 requests before and none after.
  • Still happening means there are 404 requests in both windows.
  • Unmeasured means either window lacks full status coverage, the after window is still open, original request hosts are unknown, or the before window has no observed 404s. With no 404s before, Results cannot say they stopped.
Both windows need response status codes for bot requests on every UTC date they touch, for every host in the action’s missing-page evidence. Cloudflare log pull and edge middleware provide response status; the browser snippet cannot measure 404s. With partial coverage or unknown hosts, the result stays Unmeasured and missing counts show as unavailable. Results says which condition applies: an open window, unknown hosts, missing response data, or no 404s before. To get this measurement, save the suggested redirect destination as the action’s published URL. If you save the missing URL itself, DevTune counts it as a redirect only after it observes a redirect response and no successful page on that path across the original request hosts, and it stays unidentified while the window is open. Creating a page at the missing URL does not add this measurement. Counts combine trailing-slash variants of the requested path; uppercase and lowercase paths are measured separately. These counts describe what moved. They do not change the action’s overall result or establish that the redirect caused the change. If a count is hidden while the state shows Stopped or Still happening, the count was updated after its window was measured. The state still uses the 404 requests recorded at measurement time. The same evidence is in items[].evidence.redirect404 from List measured results and devtune_get_outcomes_ledger.

Requests after a robots.txt change

For a robots.txt action you mark live, Results shows Requests from unblocked agents: for each AI provider and crawler type the change allowed, requests to every page on the sites whose robots.txt changed, in the same 14-day windows before and after.
  • Answer fetchers, such as Claude-User or ChatGPT-User, are counted as answer fetches. Search crawlers, such as OAI-SearchBot or Claude-SearchBot, are counted as crawls. The two are never added together.
  • Each line counts only the sites whose robots.txt change allowed that crawler. If one site’s change lets in Claude-User and another’s lets in OAI-SearchBot, Claude-User is counted only on the first site and OAI-SearchBot only on the second. With more than one site, each line names its sites.
  • AI Traffic records the provider and crawler type, not the exact user-agent name. Each line combines every crawler of that provider and type, so Anthropic answer fetches include Claude-User and the older Claude-Web.
  • Requests for robots.txt itself and for files that aren’t pages, such as images, are left out.
  • Counts appear once the window closes. Each site needs a server-side sensor for both full windows. www.example.com and example.com are separate sites, so a sensor on one does not count for the other.
    • With a Cloudflare log pull set up for that exact hostname and the whole site, the counts are verified.
    • With edge middleware or a Vercel log drain that reports from that hostname every day, the counts are shown with the note Seen by your edge sensor; coverage of every path on this site isn’t verified. Daily reports show the sensor was running, not that it sees every page.
    • A day with no report from the edge sensor leaves the counts unavailable. The browser snippet cannot see these requests, so a site with only the snippet shows the counts as unavailable. Partial coverage is also shown as unavailable.
The counts appear on one of the action’s published pages, the earliest still published. If you remove that page, they move to the next one, and they move back if you restore it. They describe what moved and do not change the action’s overall result. The same evidence is in items[].evidence.agentScope from List measured results and devtune_get_outcomes_ledger.

Result on an action

When you select Mark live on an action, its page gains a result section above What shipped:
  • While the window is open, the section is titled Change so far and compares the 14 days before with the days so far, showing when the window closes.
  • Once the window closes, the section is titled Result, shows the result, and compares the 14 days before with the 14 days after.
The section carries the same citation, answer-fetch, and AI referral evidence as the measured action detail in Results. If your plan does not include Results, the section says so. A spot check on the action is not part of its result. The action’s Lifecycle shows when it went live, when measuring closes, and when the result was read. Reopen moves a live action back to Accepted without discarding its measurement. Marking it live again continues the same window. See Move an action.

Set the go-live date

If you published the change before you marked the action live, set the day it went live so the before window doesn’t include days after publication.
  • When you mark an action live, choose the day under Go-live date (UTC). It defaults to today.
  • On a live action, open the menu and select Edit go-live date.
You can choose today or a day up to 14 days back. DevTune rebuilds the action’s measurement from that day: the 14 days before it and the 14 days after. A measurement whose window has closed, or one dated by a recorded page change, keeps its date. The API and MCP set the same date with liveOn on the mark_live transition.

Published pages and tracking capacity

Marking an action live, or saving a published URL, starts measurement from the date the action went live, shown as Went live. New pages and updates to existing pages do not need to wait for a crawl. Citation, answer-fetch, and referral comparisons still depend on available source coverage. Missing evidence is shown as unavailable. While your plan’s page allowance is full, a banner at the top of each project screen shows how many pages AI answers cited in the last 30 days that DevTune does not track. Select See untracked pages to list them under Insights › Audit › Crawl coverage. Select the arrow to collapse the banner to one line. If your plan’s page allowance is full, a saved published URL that DevTune does not yet track shows Tracking limit reached. Select Upgrade to track more pages to raise the allowance; your plan defines it, and upgrading increases it automatically. Saved published pages are prioritized when space becomes available; you do not need to submit the URL again. Page tracking supports audits and content-change detection; its allowance does not block outcome measurement. Upgrading does not recreate missing historical evidence. An action shows only the status of its own published URLs. If the page has not been collected, it links to crawl settings without treating a project-level crawl failure as a failure of that URL. The same status is in Get published-page tracking status and devtune_get_page_tracking.

Unclaimed changes

Crawls detect content edits whether or not an action was recorded. Unclaimed changes lists those edits, newest first, beside any measured action that claims them. Filter by URL path and use the page-size, Previous and Next controls to review the full list. When the page has a measured action completed within 14 days of the detected change, select View existing measured action to open it. The change stays unclaimed until it is associated with an action. Claim prevents a duplicate measurement for that change. Later independent edits can still be claimed. Select Claim as measured action to create a completed action dated to the detected change. Select the claimed action to open its measurement detail. Unclaim archives an unchanged manually claimed action and removes its measurement from measured totals. The action, evidence and claim history remain. Claiming it again restores the same action and measurement, including its original title and description. DevTune rejects Unclaim when the action has been edited or gained other linked work. Automatic associations cannot be unclaimed here. Older claims without reversible history also cannot be safely reversed. No action is deleted. Unclaimed-change comparisons update daily through day 14 after the detected change. Their results remain available afterward; late evidence can still correct them. Claiming an older change succeeds immediately while its longer comparisons fill in. Those comparisons show Updating until ready; missing historical evidence remains a coverage gap. The list shows 14-day before/after citation and presence differences. These differences are correlation, not proof of what caused the change. Summary totals cover the selected date range. The path filter affects only the list and its count. The same list is in List detected content changes and devtune_get_content_changes, with Claim and Unclaim as devtune_claim_content_change and devtune_unclaim_content_change. Listing needs outcomes.read; Claim and Unclaim need actions.write. Unclaim needs the current claimToken from the list, so an old request can’t undo a later claim. Agents must get human approval before changing a claim.

Next steps