Skip to content
For ad-tech, privacy and investigative researchers

Is this app sending health answers to a third party before consent, serving ads to kids under an adult flag, reporting ad impressions with no ad on the screen, syncing ad IDs after a tracking refusal, or sending an advertising ID its store label calls "No data collected"?

Ask one question across a list of apps, or record thousands of apps on real Android and iOS phones. Answers are cited to file and line. Sessions are recorded with decrypted requests, screen video and taps.

Pay as you go at $0. Questions and device hours are paid from the balance.

One wide net on real phones
sessions
2,153
apps
2,190
days
about 4
decrypted requests
770,704

Sessions of 15 minutes; a group of apps was recorded in some sessions. 19,874 ads recorded.

  • Exit by US state or country
  • Sessions up to 7 days
  • Claude Code and Codex plugin

01 / Two ways to work

Questions across a list, or a wide net on real phones.

  • Answer questions across a list of apps

    Input
    A list of apps as a CSV, package ids or store links, and the questions to ask.
    What runs
    A rule scan of each app against 833 rules, Deep scans of the targets, and the questions answered from each app's decompiled code. The list is widened or narrowed by SDK, host or identifier from Atlas, an index of 1.5 million app builds. Claims that need runtime evidence are checked in targeted sessions on real phones.
    What comes back
    Answers cited to file and line, findings per app as JSON and CSV, identifiers shared across apps, and session evidence for the claims that need it.
  • Run a wide net on real phones

    Input
    A category, an SDK cohort or a list of hundreds to thousands of apps, and the conditions: exit country or US state, persona age, language, location, tracking allowed or refused.
    What runs
    Automated sessions on real Android and iOS phones, under those conditions.
    What comes back
    Decrypted requests and HAR files, screen video synced with taps, ad records matched to the ad on screen, and search across sessions with CSV export.

02 / Sessions

What a session on a real phone records.

In a session, the app is installed on an Android or iOS phone and run under the conditions set for the study.

Conditions

  • Phones. Android phones from Pixel 6 to Pixel 9a, an iPad and an iPhone 11.
  • Exit. Exit IP in a chosen US state or country, or the researcher's own http or socks5 proxies.
  • Persona. Persona age from 5 to 99, device language, GPS location, and tracking allowed or refused. Mock Wi-Fi and cell networks on recent Android phones.
  • Length. Sessions from 60 seconds to 7 days, in segments of 5 minutes to 6 hours. One session can cover up to 20 apps.
  • Sign-ups. Sign-up flows with lab mailboxes, and web funnels opened in Chrome.

Evidence

  • Traffic. HTTPS decrypted. Requests before and after the consent prompt are listed separately.
  • Screen. Screen video, taps and requests on one clock. The ad on screen is matched to the response that served it.
  • Bodies. Request bodies of common ad and analytics SDKs decoded into named fields.
  • Live control. A researcher takes over the phone from the browser at any point. When the automated session is stuck, a help request is sent to the researcher. Each input is recorded as an attributed event.
  • Search. Search across sessions by host, field or value, with facets and CSV export.

Before consent24 requests

Network capture filtered to the requests made before consent, with the method and URL of each

Fields sent

Network capture requests with the URL of each and the fields each request sends

Screenshots from the published Temu teardown

03 / Findings

Finding types from sessions and code scans.

Each type below comes from a study run on CanITrustThat. Apps, companies and laws are named in published teardowns only.

  1. Consent timing. Health-quiz answers, email addresses and device identifiers sent to a third party, each request timed against the consent tap.

  2. Ad supply chain. For the ad on the screen: the exchange that sold the slot, the buyer, the clearing price and the child-directed flag, taken from the decrypted ad traffic.

  3. Ad events against the screen. Impressions reported while no ad was on the screen, clicks reported with no tap, a skip recorded as a click, one ad counted twice. The video and the request are shown side by side.

  4. Child and opt-out signals at scale. Child flags absent or set to adult in hundreds of the 2,190 apps of one wide net. Global Privacy Control sent while identity syncs continue. The advertising ID sent after a tracking refusal.

  5. Store label against traffic. Store listings that state "No data collected", beside a recorded request that sent the advertising ID.

  6. Sign-up funnels and offers. Sign-up flows, emailed offers and prices shown to test buyers, recorded with lab mailboxes and personas.

  7. SDK cohorts. Which apps contain a given SDK, host, key or certificate, across the app builds indexed in Atlas.

A code finding is a statement of what a build contains. A session finding is linked to the recorded request and the screen at that moment.

04 / Questions

One question, answered from the code of each app in a list.

A model reads each app's decompiled code and answers with the file and line behind each claim. Answers for a project appear in one table, one row per app. Identifiers shared by two or more apps are listed in the same table, with the apps that share each one.

How to check an answer

  • Open in Evidence

    Each citation links to the decompiled file at the cited line, with the evidence line marked.

  • Ask a follow-up

    Up to 20 turns on the same build, each answer cited the same way.

  • Scrutinize

    Select findings and queue a separate review. Each finding gets a verdict (confirmed, false positive, inconclusive or needs runtime) with its rationale and citations.

A model writes each answer. Before it is saved, each line it cites is checked against the decompiled build, and an answer that cites a missing line is rejected.

First answer in about 30 minutes. Follow-ups in about 5.

05 / Code scans

52 rule sets, Deep scans and Ultra-deep scans of the code.

Each Android app is decompiled and each iOS app decrypted, then matched against the rule pack before the first question.

  • Rules. 833 rules in 39 families, matched against the code and the store listing. 478 name a law or platform policy.
  • Rule sets. Each app in a project gets one status per rule set. The evidence still needed for a status, a session or the privacy policy text, is listed beside it.
  • Deep scan. A 7-stage AI review of the code behind network requests, ads and links. Up to 30 findings, each cited to its file.
  • Ultra-deep scan. One app read file by file, at about 10 times the tokens of a Deep scan. Hundreds to thousands of findings.
  • Native code. Native libraries and iOS binaries decompiled and added to the sources a question is answered from.

Among the 52:

  • COPPA
  • GDPR and ePrivacy Directive
  • UK GDPR and Age Appropriate Design Code
  • CCPA as amended by CPRA
  • CIPA
  • Washington My Health My Data Act
  • Video Privacy Protection Act
  • PADFA and the DOJ bulk data rule
  • China app personal information rules
  • India RBI Digital Lending Directions
  • Google Play Families policy
  • Apple App Store Review Guideline 5.1

A status is a screening result from a static scan of the package and the store listing.

06 / Plugin

Run studies from Claude Code or Codex.

With the citt command, a researcher submits app lists, starts and watches sessions, syncs raw captures to disk, searches across sessions, answers help requests, pulls Atlas cohorts and downloads exports. Prompts in the plugin are answered by the models of the researcher's own Claude Code or Codex account.

07 / Two paths

Run it on our phones, or commission the study.

  • Drive it yourself

    Plans from $800 a month with questions, scans and device hours on real phones. A device hour is one hour of session time on one phone, counted by the minute.

    • Device hours past a plan: $10 an hour from the balance
    • Dedicated fleet: phones reserved for one organization, 24 hours a day, from $1,500 per phone a month
  • We run it

    A researcher scopes the study, runs the sessions and questions at scale and writes the report: per-app and cross-app findings with screenshots, per-company reports, and each finding marked established, probable or open.

    • Question sweep: one set of questions across a list of apps
    • Wide-net investigation: automated sessions across a category or cohort, under set conditions

    Starting $12,000

Start with Pay as you go, no card.

3 rule scans a day at $0. Questions at $25 and device hours at $10 from the balance.