71 guides · PMs
Own the number before you own the tag
The expensive mistakes are product mistakes: Purchase on the landing page, no owner for GTM, staging ids in production. The validator catches syntax. You catch intent. These pages are the ones to send the room before kickoff.
Validate a pixel or CAPI All docs
Pixels and tags
A pixel is an HTTP request with a job
A tracking pixel is an HTTP request that records an ad impression, click, page view, or conversion. The transport may be a 1x1 image GET, navigator.sendBeacon, or fetch. Email open pixels and VAST Impression beacons are the same job on different surfaces.
Impression and click are not the same URL
An impression pixel fires when the creative loads and should return 200 with a 1x1 or 204. A click tracker fires on click as a 302 chain. Never 302 an impression to a landing page. View-through is a report setting, not a third pixel.
Thank-you page pixels miss the backend sale
A conversion pixel fires in the browser on the thank-you page. An MMP postback is a server-to-server GET or POST with macros expanded at conversion time. CAPI is a JSON POST to the vendor. They fail in different ways. Unexpanded macros on a fired postback are a trafficking bug.
The vendor host cannot see your cookie jar
Third-party cookies on facebook.com or google-analytics.com are blocked or partitioned in Safari, Firefox, and Chrome. First-party pixels set _fbp and _ga on your domain. CHIPS partitions third-party cookies by top-level site. ITP expires tracker-like storage. Matching moved to click ids and server events.
The URL changed. The pixel did not.
Single-page apps do not reload the document. If you only fire PageView on first load, checkout and thank-you routes are invisible. Hook history. React 18 Strict Mode double-mounts and will double Purchase. Meta CAPI event_source_url must be the current route, not the first URL of the session.
UTM is not a click id
utm_source, utm_medium, utm_campaign, utm_content, and utm_term are for your analytics property. Google Ads, Meta, and TikTok attribute paid clicks with gclid, fbclid, and ttclid. UTM is not a click id. Casing splits sessions. Do not put gclid in utm_content and expect Ads to see it.
Purchase does not belong on the landing page
PageView belongs on the landing page, where click ids are stored. Purchase belongs on the thank-you page, once. Hosted checkout (Shopify, Stripe Checkout) often means the thank-you page is another host, so cookies and click ids must travel or the conversion fires as a new user. Firing Purchase on every landing inflates ROAS until the platform discounts you.
Not every 200 was a person
Prefetch, Slack unfurls, Gmail image proxies, and link scanners fire pixels without a user. Platforms filter invalid traffic in the UI. Your BigQuery export of raw hits will not match. Gate conversion tags on a real order id, not on DOM Ready of a public URL.
The tag ran. The request did not.
Ad blockers drop known pixel hosts from EasyList and similar lists. GTM preview can say the tag fired while facebook.com/tr never leaves the machine. Noscript does not bypass a host block. Recovery is CAPI or sGTM on a first-party collector you operate, not a second request to the same listed host.
buyer@example.com in the query is a finding
Raw emails, phones, and names in a pixel URL land in vendor logs, CDN logs, and Referer headers. Meta flags unhashed emails on facebook.com/tr as vendor.meta.pii.unhashed_email. Hash SHA-256 of a normalized email in a body field, or use CAPI. Image pixels cannot hash in the browser.
A first-party hostname is still their collector
Serving a vendor collector from metrics.example.com via DNS CNAME can dodge some blocklists. Safari classified CNAME-cloaked trackers. Certificate automation and subdomain takeovers become your incident. Server-side GTM on infrastructure you own is a different design: you send the event, you do not pretend facebook.com is you.
Conversion APIs
One conversion API JSON cannot serve every vendor
Meta Conversions API event_time is seconds. Reddit CAPI event_at is milliseconds. LinkedIn conversionHappenedAt is milliseconds and still needs user.userIds. TikTok timestamp is ISO 8601. A shared conversion API event named event_time will be wrong on at least one hop.
What is a conversion API
A conversion API is a server POST that sends conversion API events your backend already knows: purchases, leads, app events. Meta names it the Conversions API (CAPI). TikTok names it Events API. Google names it Measurement Protocol. Same job, incompatible JSON.
Two pipes, one conversion
The recommended setup for Meta, TikTok, and others is browser pixel plus conversion API events, deduplicated on event_id. Drop either side without a plan and you undercount or double-count.
One event_id, two transports
Meta, TikTok, and others drop a duplicate conversion when the same event_id arrives on the pixel and the conversion API. Generate it once, persist it, send it on both pipes.
A conversion with nobody on it
Match quality is how confidently the platform attaches your event to a user. Click ids score high. Hashed email and phone help. IP plus user-agent are weak but real. Empty user data is a dark event.
sGTM is a proxy. CAPI is your job.
Server-side Google Tag Manager is a first-party collector that forwards to vendors. Direct CAPI from your backend skips the tag manager. Both are server events. They are not the same architecture.
200 means the server was awake
PostHog capture, Segment track, and several ads CAPIs return 200 for payloads they will not ingest. Validate the body before you ship. Do not use status code as the contract.
The conversion happened in Salesforce
Offline conversions upload events that never happened in the browser: phone sales, in-store POS, Salesforce stages. Matching is PII-heavy. Timestamps are usually delayed. Dedup against any pixel you already fired.
Purchse is a custom event whether you meant it
Pinterest standard events are lowercase (pagevisit, checkout, addtocart). Meta is PascalCase (Purchase, Lead, AddToCart). Snap is UPPER_CASE. A typo becomes a custom event and still 200s.
Do not fire the whole funnel on thank-you
Funnel events exist so the platform can optimize toward the next step. Firing all of them on the thank-you page teaches the model that every user is a buyer. Firing none of them except Purchase starves upper-funnel learning.
A refund is a measurement event
If you send Purchase and never send Refund, the ads platform keeps spending toward orders that came back. Meta, GA4, and several CAPIs have refund events. Use the original transaction id.
Check the CAPI body before Test Events
A conversion API validator checks CAPI JSON against the vendor contract: Meta event_time in seconds, hashed em, action_source, event_source_url, Purchase value and currency. Paste the body. HTTP 200 is not that check.
Identity and matching
gclid, fbclid, ttclid, msclkid, li_fat_id
Paid platforms join conversions on click ids: gclid, fbclid, ttclid, msclkid, li_fat_id, twclid. UTMs do not replace them. Capture on the landing URL and send them again on the conversion.
client_id is analytics. external_id is yours.
GA4 client_id is the analytics browser id. Meta external_id is your customer key, hashed. Firebase app_instance_id is the app stream id. They are not aliases. Do not paste one into the other field.
distinct_id is who PostHog and Mixpanel join on
PostHog requires distinct_id on every capture, including each batch row. Mixpanel joins on properties.distinct_id. Segment uses userId or anonymousId, which destinations often map to distinct_id.
The platform's graph is not your CDP
Meta, Google, and the CDPs each keep a graph you do not own. You send hashed PII, click ids, and cookies. You do not download their device graph. Matching is a payload problem, not a warehouse join you can copy.
Safari will expire that cookie
Safari Intelligent Tracking Prevention blocks third-party cookies and caps first-party cookies that look like trackers. _fbp and _ga will expire sooner than your tag's max-age. Design for click ids and server events. CNAME cloaking is classified.
Checkout on another host drops the click id
Checkout on another host drops first-party cookies and often the click id. Shop.example.com does not see _fbp from www.example.com unless Domain=.example.com. Forward gclid, fbclid, and the cookies you need, or fire CAPI from the order. GA4 linker uses _gl.
Time and attribution
Ingestion time is not event time
Ingestion time is when the collector got the HTTP request. Event time is when the purchase or click happened. Queues, retries, and epoch timestamps on ISO fields collapse the two. Send the time of the action, not Date.now() at drain.
The window is a report setting, not a pixel
The attribution window is a report and bidding setting on the ads platform, not a pixel parameter. Changing it rewrites history in the UI. It does not change what the tag sent. View-through and click-through are those settings, not two pixels.
Too late for the graph, still true in finance
Ads graphs drop conversions that arrive after a maximum age. Meta CAPI is about 7 days. LinkedIn CAPI is 90 days. GA4 MP backdates about 72 hours. Finance still needs the late sale. Send it where the vendor allows, and keep a warehouse ROAS the ads UI will not match.
View-through is an impression credit, not a click
Click-through credits a click. View-through credits an impression with no click, inside a (usually shorter) window. They are report settings, not two conversion tags. Mixing them in a ROAS target double-counts attention.
Last click is a story. Data-driven is a model.
Last-click gives the conversion to the final paid click. Data-driven spreads credit across touches the platform observed. UTM last-click in GA4 is a third story. None of these is the pixel. They are models on top of the hits. Windows still apply.
Consent and privacy
A TC String is not the letters 1 or true
gdpr must be 0 or 1. gdpr=1 needs a TC String in gdpr_consent. Pixellint decodes the string. gdpr_consent=1 and gdpr_consent=true are valid base64 and still not a TC String. Core reads query string and Floodlight-style path parameters.
GPP header type is 3
A GPP string has header type 3. A TC String pasted into gpp decodes as type 2. gpp needs gpp_sid so the callee knows which section is in force. The spec does not require sid to match sections inside the string. Pixellint does not invent that check.
US Privacy was deprecated on 31 January 2024
The IAB US Privacy string is a version digit plus three Y, N, or hyphen characters. Only version 1 was published. IAB Tech Lab deprecated it on 31 January 2024 in favor of GPP. Pixellint still checks the shape and always warns that the parameter is deprecated.
Consent Mode is a gtag signal, not a CMP
Google Consent Mode is a gtag or GTM signal for ad_storage, analytics_storage, ad_user_data, and ad_personalization. It is not a CMP and not a TC String. Default denied, then update on choice. EEA traffic without the v2 keys is a policy problem, not a pixel syntax problem Pixellint invented.
A banner that does not block still fires the pixel
A banner that does not wrap tags still fires pixels on first paint. Blocking means the tag does not queue until granted, or Consent Mode defaults to denied. UX copy is not a technical control. Network tab on a fresh profile is the audit, not a screenshot of the modal.
CAPI still needs a lawful basis
Moving the pixel to CAPI does not move you out of GDPR. You still need a lawful basis, a processor story, and a way to respect refusal. Hashing an email is not anonymization. The vendor 200 is not a legal opinion. Consent Mode does not update your shop backend.
Limited Data Use is a parameter, not a vibe
Meta Limited Data Use is dpo=LDU plus a country and usually a state. dpoco=1 is the US. Country without state lets Meta geolocate. Pixellint flags that as vendor.meta.ldu.country_without_state. On CAPI, data_processing_options_country is required when LDU is sent.
ATT is a prompt. IDFA is the identifier.
ATT is the iOS prompt. IDFA is the advertising identifier you may read after authorize. Denied ATT does not forbid first-party CAPI with hashed email you collected in-app. It does forbid IDFA. All zeros means denied. Do not hash IDFA like email unless the MMP field is documented as hashed.
Child-directed traffic is not a tag setting you forget
Child-directed apps and mixed-audience properties cannot treat ads pixels as default. Flag child-directed traffic in the vendor's restricted data processing settings. Hashing an email you should not have does not help. Pixellint does not invent a COPPA rulepack. The vendor 200 is not permission.
Event logs are PII adjacent
Pixel query strings, CAPI bodies, HARs, Segment archives, and sGTM logs are identifier stores. Keep them as short as the debug job requires. A six-month access log of hashed emails is still a dossier. Pixellint flags a raw email on a live artifact. It will not scrub the file you already uploaded.
Cookies are not the only measurement future
Third-party cookies are already gone in Safari and Firefox. Chrome moved to user choice and Privacy Sandbox APIs. Topics and Attribution Reporting are not a pixel with a different host. Keep CAPI while you learn them. Privacy Sandbox does not waive GDPR or a CMP. Pixellint does not validate Topics values.
Aggregate reports are not a pixel HAR
The Attribution Reporting API sends delayed, noisy aggregate reports (and limited event-level reports) from the browser. It is not a conversion pixel you can paste into a HAR validator. Keep enhanced conversions and gclid. Pixellint does not validate Attribution-Reporting-Register-Source headers.
Withdrawing consent has to hit the same URL that accepted it
Granting consent is a request the vendor can see. Revoking it often is not. A banner that shows a new choice while Floodlight, Meta, and CAPI keep the old string is a withdraw that never left the page. Lint the hop after refuse, not only after accept.
Mobile and MMP
The MMP endpoint is a contract, not fire-and-forget
Adjust, AppsFlyer, and Branch S2S endpoints are contracts: required tokens, device ids, clocks, and unhashed IP. A 200 from the MMP does not mean the event was recorded. Validate the body before you retry. Pixellint packs encode the documented fields, not SKAN postbacks.
SKAN is a conversion value, not a click id
SKAdNetwork attributes iOS campaigns with a conversion value and a delayed Apple postback. It is not a click id, not an MMP device id, and not a replacement for S2S in-app events. Map values on purpose or the postback is a coin flip. Pixellint does not validate SKAN postbacks.
Install is not session is not in-app purchase
Install, first open, session, and in-app purchase are different MMP events. Firing purchase on first open inflates ROAS. Firing install on every session inflates CPI. Name the event the vendor catalog uses, then S2S it once. Adjust tokens are opaque. AppsFlyer eventName is a string. Branch distinguishes standard and custom endpoints.
The click was on the web. The conversion was in the app.
A paid click on the web that converts in the app needs a bridge: MMP, Meta app events, or SKAN. The web pixel alone stops at the store page. Store the web click id before the user leaves for the store. action_source=app on CAPI. Do not fire website Purchase when the charge happened in IAP.
IDFA and GAID are advertising ids. IDFV is not.
IDFA vs GAID: IDFA is iOS advertising, gated by ATT. GAID is Android advertising (gps_adid, aaid, advertising_id). IDFV is per-vendor on iOS, not an ads cookie. Sending the wrong one in the wrong MMP field looks populated and matches nobody. Zeros means denied.
Store receipts are the source of truth
Store receipts are the source of truth for IAP. SDK-observed purchases can fire twice, fire on restore, or fire before charge. Validate with Apple or Google, then S2S to the MMP with the real value and ISO currency. Restores are not purchases. A 200 from the MMP does not prove the revenue dashboard moved.
Device graphs got worse. Probabilistic got noisier.
Probabilistic matching (IP, user-agent, time) filled gaps when IDFA was everywhere. After ATT it is noisier. SKAN, consented IDFA, and deterministic MMP clicks are the planned replacements, not a better fingerprint. Do not hash IP. Pixellint flags hashed IP on Adjust and AppsFlyer. It will not tell you the unhashed IP was CGNAT.
Analytics and CDP
timestamp_micros is 16 digits
GA4 MP needs api_secret and exactly one of measurement_id (G- form) or firebase_app_id. timestamp_micros is 16 digits. Purchase needs currency, value, transaction_id, and items. HTTP 204 is not a schema review.
PostHog returns 200 and drops the event
PostHog documents that a capture missing event or distinct_id still returns 200 and is not ingested. Send ISO 8601 for event time. Do not treat status code as a contract test.
track without an event name still 200s
Segment's HTTP Tracking API track call needs an event name. The HTTP 200 does not check it. anonymousId or userId still have to exist if you want a person, not a floating event.
Mixpanel join key is distinct_id
Mixpanel joins on distinct_id. Engage/identify must reuse it. Time is epoch seconds in some Mixpanel APIs and not others; read the endpoint you POST to, not a blog from 2016.
A CDP is a pipe. A pixel is a bid.
A CDP (Segment and friends) routes events. An ads pixel or CAPI is a bid and attribution pipe. Product analytics is product. One track() call is not automatically all three jobs done well.
Sampling hides the broken 1%
GA4 UI sampling and Mixpanel bookkeeping hide rare broken payloads. QA on 20 events will never see the 1% of Purchases missing currency. Sample raw exports or contract-test fixtures.
Product analytics and ads conversion are different jobs
Product analytics wants funnels and retention. Ads pixels want matchable conversions for bidding. Sharing one Purchase event is fine. Sharing one schema without a map is how both teams think the other is wrong.
Launch and QA
Ship the dictionary before you ship the tag
Ship the event dictionary, owners, environments, PII rules, and the success metric before anyone pastes a snippet. A tag without a dictionary is how Purchse reaches production.
Test Events is empty until both pipes exist
Before go-live: both pipes, shared event_id, Test Events showing matched user data, value and currency on money events, and a plan for refunds. Empty Test Events means you are not live, even if GTM says the tag fired.
Staging must not train production
Staging must not train production. Use test_event_code, debug_mode, or a separate pixel / measurement id. Pointing staging at the production pixel because 'it is easier' is a data incident.
Fixtures in CI, not a weekly HAR
Keep fixture URLs and JSON next to the code. pixellint validate url and pixellint validate json on every PR. Weekly HAR archaeology does not catch the merge that shipped Date.now() to Meta.
A HAR is a fired artifact plus PII
A HAR is a fired artifact plus cookies plus whatever PII the page put in query strings. Share it like a password dump. Prefer staging HARs and redaction.
GTM said fired. Network is empty.
Facebook pixel not firing, TikTok pixel not tracking, GTM says the tag fired and Network is empty. Check the collector URL, ad blockers, consent, SPA routes, duplicate installs, and GTM environments. Then paste the URL into a pixel validator.
Vendor playbooks
OpenAI CAPI is not Meta with a new host
bzr.openai.com/v1/events wants events[].id, type, and timestamp_ms as 13-digit milliseconds. Meta wants event_name, event_time as 10-digit seconds, and event_id. Copying a Graph API body to OpenAI is how the event never attributes.
fbq plus CAPI, same event_id
fbq in the browser plus CAPI on the server, same event_name and event_id. event_time in seconds. Hash em and ph. Send IP and UA in the clear. _fbp and _fbc when you have them.
Insight Tag plus CAPI in milliseconds
The Insight Tag is often an image pixel. CAPI is JSON with timestamps in milliseconds. Lead gen on LinkedIn plus a CRM upload is a third path. Do not mix the clocks.
Conversion linker, gclid, then the conversion tag
Conversion linker writes first-party cookies. The conversion tag fires on thank-you with send_to. Enhanced conversions add hashed email or phone. gclid still matters. Consent Mode still matters.