First, put the conversions beside the business's own records for the same days
The conversion troubleshooting collection sets the order for every symptom: measurement, then traffic, then the page. A conversion rate (CVRShare of visits or clicks that convert.) is a ratio, and its top and bottom can each shift while every visitor behaves exactly as before. The rule applies to a rise as much as to the fall diagnosed in conversion rate dropped overnight, and a rise is easier to let pass, because nobody complains about good news.
Start with the system the business treats as its source of truthThe system whose figure is treated as authoritative for a given number, such as billing for revenue or the CRM for pipeline, with other tools reconciled to it. for the outcome: the order system for purchases, the CRM for enquiries, the booking system for bookings. Set the analytics count beside the business count for closed days on each side of the rise. For a shop, match a single closed day order against order using the transaction identifier, as the diagnosis for GA4 revenue not matching orders does. Then read the visits behind the rate, and the rate by source, device and new against returning visitors.
If analytics rose and the records did not, look at duplicates, key events and spam. If the records held while visits fell, the denominator shrank. If the records rose too, the change is in who arrived or in the site and its offer.
Two cautions come first. A control chart of the rate, with limits drawn from its own history, shows whether the rise is a signal or the usual wobble. And the guide to reading small numbers shows why a rate resting on only a few conversions swings so much; the conversion rate calculator turns reconciled counts back into a rate to set beside the reported one.
Diagnosis table
Diagnosis
| Symptom | Check | Likely cause | Fix |
|---|---|---|---|
| More purchases or leads in analytics than the business recorded, some appearing twice | One closed day reconciled by transaction identifier or enquiry; repeats after a refresh, back navigation or an emailed link | Conversion sent again when the confirmation page loads again | Send it once from the order data with the transaction identifier; count leads once per session |
| Conversions roughly doubled from a release or container publish date; users barely changed | Preview mode and the network panel during a conversion; the template code for a second copy | A second tag or a duplicated trigger | Remove the extra trigger or hard-coded snippet |
| Key events jumped while orders and leads did not; a new name in the key events report | The key events list and counting methods against the measurement plan | A non-outcome marked as a key event, or a counting method changed | Unmark it; read the rate on the named outcome |
| Leads rose but sales rejects many as spam; nonsense, repeats, odd hours | Analytics leads against accepted CRM leads by form, source and hour; the form's protection | Automated or spam submissions accepted | Put the hidden field or a server check back, keeping a route that works for everyone |
| Sessions fell on the day the rate rose; orders or leads held | Sessions by template and day; consent, tag and filter changes around the date | Visits undercounted while outcomes are still counted | Restore the tag; log consent and filter changes as breaks |
| Records rose with the rate; one source, device or visitor type grew | Rate and share of sessions by segment, before and after | A shift in who arrives | Report by segment; credit the campaign, not the page |
| Records rose, the mix held and segment rates rose together | The change log for releases, winners, prices and promotions | A real change to the page or the offer | Confirm it against a written prediction; watch margin |
Fix, verify and prevent each cause
A confirmation page that sends the conversion again
The purchase or lead tag usually lives on the page after the purchase, the form or the booking, and every load of that page can send the conversion again: a refresh, going back to it through the browser history, a bookmark or a confirmation link reopened from an order email. The guide to confirmation pages and thank-you pages takes its rule from the key events guide: a Key EventA GA4 event marked as important to the business, the property's measure of success. fires once for each real occurrence, never because the page was refreshed, reached with the back button or loaded by a test order.
Fix: build the purchase event from the stored order, not from the page load, and include that order's transaction identifier; the confirmation guide notes that the GA4 e-commerce developer guide, listed in this page's sources, attaches a transaction identifier to every order's purchase event, which stops one purchase being recorded twice. Count leads once per session, since reloading a thank-you page does not make a second enquiry. The glossary covers DeduplicationRemoving double counts when one conversion is reported by more than one source. across systems.
Verify: reconcile a fresh closed day by transaction identifier or enquiry and find no repeats, and confirm in preview mode that reloading the confirmation page sends nothing new.
Prevent: add the reload and back-navigation check to every release that touches the confirmation page, and keep the daily reconciliation running.
A second tag or a duplicated trigger
A release or a container change can leave the same conversion sent twice: a second trigger on a tag, a new tag beside the old one, or a hard-coded snippet in the template alongside the container. The tutorial on debugging with preview mode lists two causes of a tag firing twice, a tag with two triggers or a hard-coded copy of it in the page, and looks for it in the summary view and the network panel. The signature in reports is a count that steps up on one date while users barely change, which the bot traffic diagnosis also describes for a second page tag.
Fix: remove the extra trigger or tag in the container, or take the hard-coded copy out of the template; the comparison of Tag Manager and hard-coded tags calls having both the standard double count and says to pick one route and document it. For a lead form moved to a success push, the diagnosis for form submissions not tracking deletes the previous form or click trigger in that same workspace, so no submission is counted twice.
Verify: one request per conversion in preview, and a closed day where analytics and the business agree at their usual ratio.
Prevent: walk the tracking verification checklist after every container publish and every template release that touches a tracked page.
An event newly treated as an outcome
Google's page about events, listed in the sources, describes the event as the unit analytics measures, and the guide to GA4 key events explains that marking an event alters what reports do with it, not what is collected. Once a brochure download or a form start is marked, it counts as a key event in every report, and a rate read across all key events rises from that day. Switching a lead event from once per session to once per event raises the count too, although the guide notes that the session key event rate is unaffected by the counting method, so which rate the dashboard uses decides whether that change shows.
Fix: unmark any event that is not an outcome, read the rate on the named primary outcome, and set the counting method the guide recommends: once per event for purchases, once per session for leads. Both changes apply going forward only, so annotate the affected days. Where the key event is imported into an advertising platform and shapes bidding, make the change through the conversion goal change checklist.
Verify: the key events report lists only outcomes, and the rate on the primary outcome sits back in line with the business's records.
Prevent: give the key events list one owner and record every mark, unmark and method change as a dated entry.
Automated or spam submissions counted as leads
Automated visits usually lower a rate, because they add sessions that never convert. Automated form submissions do the opposite: each one the form accepts is a lead in analytics, while the sales team sees nonsense. The diagnosis for low-quality Google Ads leads separates automated entries, which bunch in one campaign type or network, arrive at odd hours and carry gibberish, from genuine people who are simply poor fits.
Fix: restore or add protection as the forms guide describes it: a hidden field, a check on the server, or a provider that offers an accessible route, never a puzzle that excludes people who cannot see it. Exclude the entries already received from lead reports and from anything sent to advertising platforms.
Verify: analytics leads match the enquiries sales accepts, and real people still complete the form at their usual rate; if completion falls after the fix, the diagnosis for forms that are started and never sent covers what the change may have broken.
Prevent: check the form's protection in every release that touches the form, and compare analytics leads with accepted CRM leads each week.
Visits undercounted while outcomes are still counted
A rate also rises when visits leave the denominator and conversions do not. A consent change can do it: the guide to privacy, consent and measurement explains, from Google's documentation on consent modeA Google tag feature that adjusts what tags collect according to the visitor's consent choices. listed in the sources, that where a visitor declines, some observations are never made and part of what is missing is estimated by modelling, and that a rate worked out from consenting visitors alone tends to overstate the true one, since those visitors are usually the more engaged. So can a page tag lost on some templates, leaving visits there unrecorded while the confirmation page still sends its conversion, or a new exclusion such as an internal trafficVisits from the business's own offices, staff, agencies and testers, labelled by an address rule and filtered so they do not count as customers. filter, which drops matching visits from the day it is switched on.
Fix: restore the tag on every template, confirmed in preview mode; for a consent change or a new filter, record the date as a break in trend, compare periods on the same basis and say which numbers include modelling, as the privacy guide asks.
Verify: sessions by template return to their usual shape, and the rate is compared only across periods on the same basis.
Prevent: check sessions by template after every release, and log consent and filter changes with their dates.
A shift in who arrives
Sometimes the business's own records rise with the rate and the site has nothing to do with it. An email send to existing customers, a brand campaign or a cut in prospecting spend leaves a larger share of visitors who came intending to buy, and the blended rate rises even if no group converts any better. The glossary calls it a mix shiftA change in a blended metric caused by a change in the mix of segments, such as devices or sources, rather than by any segment performing differently., and the guide to explaining a change in a marketing metric separates a change in which segments make up the traffic from a change in how each segment converts.
Fix: decompose the change by source, device and new against returning visitors, as the guide sets out; report the segment rates beside the blended one and credit the send or the campaign, not the page. A rise after prospecting was cut is a finding about spend, not about the site.
Verify: each segment's own rate matches its level before the rise, and the share of sessions explains the blended change.
Prevent: annotate every send and campaign launch or pause, and read the rate by segment as standard, not only as one total.
A real change to the page or the offer
A rise that survives every check above is real: a fix that removed a fault, a winning test shipped, a price cut or a promotion. The tutorial on setting up a measurement change log has every investigation begin with the log and suggests adding trading events such as price changes when the log also serves people explaining performance, so the cause is usually a dated entry. A movement with no entry means either the business really changed or someone failed to log it, and that is settled before anything is reported.
Fix: nothing needs repairing, but the credit needs care. Read a shipped change the way the guide to improving conversion without A/B tests describes, against a prediction written in advance and comparable periods, rather than crediting the whole rise to whatever shipped last. For a price cut or a promotion, read revenue and order value beside the rate, because more orders at a lower margin is a different result.
Verify: the rise holds in the business's records over full weeks and across segments, and the explanation fits data it did not come from.
Prevent: log every release, shipped winner, price change and promotion with its date, so that the next rise has its explanation waiting.
A prevention routine: reconcile, log and alert on rises as well as falls
Most monitoring is built to catch falls. The routine that catches a false rise is the same one facing both ways.
- Reconcile every week. Set the analytics count of each primary outcome beside the business's own count for the same closed days: orders by transaction identifier, enquiries accepted in the CRM, bookings in the booking system. Alert when the ratio moves in either direction, not only when analytics falls short.
- Keep one dated change log. Container versions, releases, consent changes, filters, key event marks and counting methods, campaigns, email sends, prices and promotions. A rise with no entry beside it comes first.
- Chart the rate with limits on both sides. A control chart drawn from the metric's own history gives an upper limit as well as a lower one, so a rise outside the ordinary range raises the same question as a fall.
- Report the rate on named outcomes. Show the rate for the primary key event in dashboards, not the rate for all key events, so a newly marked event cannot inflate it unnoticed.
- Read by segment as standard. Rates by source, device and visitor type sit beside the blended rate, so a mix shift shows as what it is.
- Walk the tracking after releases. Every release that touches a confirmation page, a form or a tracked template ends with the verification checklist and a check of the form's protection.
- Confirm before crediting. No rise is reported as a win until the reconciliation, the visits, the mix and the change log have been checked, in that order.
A kitchen fitter's enquiry rate leaps (illustrative)
Gaskivonde Kitchens checks a rise before celebrating it
Gaskivonde Kitchens, an invented kitchen design and fitting company, normally records about 65 website enquiries a week from around 6,000 sessions, an enquiry rate near 1.1%. One week analytics reports 179 key events from 6,300 sessions, a rate of 2.8%, far above the upper limit of its control chart, and nobody can name a campaign or page change behind it. The CRM holds 100 form entries for the same days, 21 of them rejected by sales as spam.
| Part of the reported figure | Key events | How it was found |
|---|---|---|
| Genuine enquiries | 79 | CRM entries accepted by sales |
| Genuine enquiries counted a second time | 52 | Preview mode: a Tuesday release added a form-success tag on the thank-you page while the old form-submit tag still fired |
| Brochure downloads | 27 | The key events list: a colleague had marked the download, and the dashboard read all key events |
| Automated submissions | 21 | Nonsense in the fields, overnight, after the same release removed the form's hidden field; the bots never loaded the thank-you page, so only the old form-submit tag counted them |
The genuine rise from about 65 to 79 enquiries came from outside the site: a brand campaign lifted returning visitors from 22% to 31% of sessions, while returning visitors converted at about 2.55% and new visitors at about 0.67% in both weeks, a mix effect taking the true rate from 1.1% to 1.25%. The company removes the old trigger, unmarks the download, restores the hidden field with a server-side check, logs each change with its date and adds a weekly reconciliation against accepted CRM enquiries that alerts when the gap moves either way.
Common mistakes
- Reporting the rise before reconciling it. A rate the business's records do not share is a counting fault.
- Checking tracking only when numbers fall. Duplicate tags and spam entries make numbers better, not worse.
- Reading the rate across all key events. One newly marked event changes it without anyone noticing.
- Ignoring the denominator. Fewer recorded visits raise a rate as surely as more conversions do.
- Crediting the page for a change in who arrived. Read segment rates before the blended one.
- Crediting whatever shipped last. A real rise is read against a prediction and comparable periods.
