What our ads actually did

Detected The ad body was 563 characters. Meta shows about 125 before a "See more" that roughly 1% of readers tap, so the part a phone displayed cut mid-word inside the problem statement and never reached what the product does. The image said "aren't working", the ad headline said "Ask your client's site why", and the page H1 said the same as the headline: three phrasings of one idea. The Slack card on the image carried "62% of paid visitors" and "214 sessions" with nothing marking it as an example, on an ad whose whole promise is not publishing numbers we did not measure. And the image never showed Slack itself, only the word.

Changed Body cut to 122 characters so all of it shows above the fold. Image, ad headline and page H1 are now one sentence. The Slack card is labelled EXAMPLE, drawn with the same Slack sidebar chrome the homepage uses, and carries the Slack mark plus the four assistant logos the page shows. The creative is generated from the brand tokens by scripts/build-ad-creatives.js instead of by hand, so it cannot drift off-brand again, and create_adcreative now runs the copy validator before the API call, which is how 563 characters got through in the first place. Ad AGT_Harvv_Agencies_AskWhy_2026-09-06_ad3_slack replaces ad2; nothing else in the ad set moved.

Detected At 384x696, the most common screen in this campaign, the consent bar covered the bottom third of the first screen: exactly where the Slack example that the ad promises begins. The pixel's reading signal showed the subtitle was the only block most phone visitors ever had in view: 32 saw it and skipped, 7 skimmed, 11 read; every block below it was never in view for 30 or more sessions. 92% of phone visitors never scroll, so they saw the promise and not the proof. The most tapped thing on the page was "Accept all".

Changed Banner copy cut from 22 to 11 words and its mobile padding tightened. Subtitle cut from 31 to 15 words and the hero pulled up. Rendered at 384x696, the question and Harvv's answer are now both on the first screen. Not changed: the banner still shows, because UK visitors are in the targeting and a returning-visitor cookie needs consent there.

Detected On a real Android phone the URL field on /ask measured about 390px tall and the email field the same; on /ads-audit both were 190px. The rule was flex: 1 1 190px on the inputs, a sensible width in the desktop row, and on phones the form flips to a column, where flex-basis becomes the height. Four layout checks passed while the field we ask a paid visitor to type into was taller than their viewport is wide. /ads-audit was also rendering in the dark theme against a light-default site.

Changed flex-basis reset to auto inside the column rule on the three pages whose form flips (/ask, /ads-audit, /audit); /ads-audit rendered light. The layout gate now fails any text input taller than 80px at phone widths, and a planted copy of the bug reddened it before this shipped.

Detected Across 116 ad sessions in 5 days the median ACTIVE time on the page was 1.1 seconds, and 97 of them ended under 3 seconds. The wall-clock median was 31 seconds, which is a tab left open in the background, and reading that as attention is what made the traffic look ordinary. 107 of 116 never scrolled. Nothing on the page was touched: no form, no calendar click, no audit submit.

Changed The page now fires an "engaged" signal at 15 seconds of active time or 50% scroll, whichever comes first, and /experiment counts it as a branch beside the calendar and the audit. It is not a conversion and is never drawn as one. It exists to separate an ad that bought a reader from an ad that bought a page load, which no number already on this page could do. It is also the only optimisation goal this budget can reach: Meta needs about 50 events a week to learn, and at a $75 target cost that would be $536 a day against the $10 we spend.

Detected The ad sells an answer arriving in Slack or in your AI assistant. The page opened with a headline, a question, a paragraph, two buttons and a form asking for a URL and an email, and the Slack answer panel sat below all of it. Read-depth confirms it: the only blocks any visitor ever saw were the opening paragraph and the form subtitle, both above the fold, and every deeper block was seen only by the handful of sessions that scrolled at all.

Changed The Slack answer panel moved up into the first screen, directly under the opening line, with the question chips beneath it. The buttons and the audit form moved below it. The headline now matches the ad word for word, and the opening line names Slack and the assistant instead of describing the idea abstractly. Nothing was added or removed, only reordered and reworded.

Detected The ad set was set to the United States, Canada and Australia. Meta delivered 3,299 of 3,412 impressions to Canada. The United States got 16. Asking for the cheapest possible page load, which is what the optimisation goal and the bidding strategy together ask for, gets the cheapest country, and we never tested the market we meant to.

Changed Canada is out of the ad set and the United Kingdom is in, so it is now US, GB and AU. Everything else is untouched: the same interests, the same placements, the same budget. Canada is not being re-measured because 116 sessions of it is already enough to know. This will cost more per click and buy fewer visits a day, and that is the trade being made deliberately.

Detected Meta was selling us phones almost exclusively: 55% Android smartphone, 39% iPhone, 2% tablet, across 5,983 impressions. Our own pixel says phones do not read this page. On /ask over 7 days, excluding flagged traffic: 123 mobile sessions, 92% of which never scrolled at all, median active time 1.17 seconds, 2 that reached halfway. Against that, 9 desktop sessions, 56% never scrolled, median 3.77 seconds, 3 that reached halfway. Three of nine against two of 123. Desktop is a small sample and is reported as a lead rather than a proof, but site-wide the same split holds at larger n: desktop median active time is 172 seconds against 4 to 18 on mobile. This is not the scroll blind spot; every unload event carries the scroll high-water mark, 318 of 318 on mobile, and it reports zero.

Changed The ad set now asks Meta for desktop only. Estimated reach falls from about 138,500-163,000 monthly actives to 50,800-59,800, which is still far more than a $10 a day budget can exhaust. Nothing else moved: still Facebook feed only, still US/GB/AU. This is the cheapest test available of the one difference the data actually shows, and it is one setting to undo. Read the before-and-after on this page with care. Two audience changes landed within an hour of each other today: Canada was dropped for GB earlier, and Canada had been 3,299 of 3,412 impressions. Every rate that spans 2026-09-05 therefore compares three different audiences, not one, and no verdict should be drawn across that boundary.

Detected The funnel showed 60 landing page views from Meta and 17 sessions from our pixel, with a fixed sentence underneath saying our count was the larger. It was written when that was true and never re-checked against the numbers it sat under. Nothing on the page counted what happened between the two meters.

Changed The sentence is now computed from the numbers, and a new block counts each step: our server answering the request, the pixel reporting a session, the bot classifier keeping it, and the no-click gate keeping it. Every drop between two adjacent counts is one named cause. The change log states that figures quoted inside an entry are as of that date.

Detected The goal block divided spend by 17 sessions and printed $1.25 a session, and the trend strip on the same page refused the same division as too thin. The thresholds derived from that cost had tripled in two days. Two rules for one quotient on one page.

Changed One floor, defined once in the analysis module and used by both. Under 25 sessions the goal block says what it is waiting for instead of printing a cost.

Detected The first day of the campaign read 13 impressions, 0 clicks, 0 landing page views and 7 sessions. The ad-review crawlers Meta sends at ad creation arrived on the same calendar day the campaign started spending, so the day-level spend gate let them through as visitors.

Changed A day with spend but no Meta link click did not deliver a visitor from this ad. Tagged sessions on such a day are zeroed with a note, and the raw count stays visible in the reconciliation block so the removal is not silent.

Detected The trend strip labelled a 2.55% click-through rate and a $0.34 cost per click "best in class" against published ranges, on a campaign where no session had reached the signup page. Both figures were true and the word was praise for a stage whose output had not been measured.

Changed The standings now read "top of the published range" rather than "best in class", and every upstream metric carries the share of its sessions that read past the fold, or says that share is not yet measurable. A new metric, cost per session that read past the fold, sits beside cost per session.

Detected Measured in a real browser at phone size: the consent banner was 281px tall and covered a third of the first screen, sitting on top of both secondary buttons. Over the same 30 days its reject button was clicked more often than the page own audit button.

Changed Rebuilt it as a compact bar. It now covers 18% of the first screen instead of 33%, and its buttons meet the 44px minimum tap size we tell customers to meet.

Detected The chat button and the live tracker dock were pinned to the same corner with the same offsets. The dock renders above everything, so it covered 59% to 77% of the chat button at every scroll position. The button had zero clicks in 30 days.

Changed Moved the chat button to the opposite corner, and above the dock on phones where the two do not fit side by side. Nothing now covers it.

Detected On phones the three cards explaining what we find were collapsed, so the button inside each one was never drawn. Those buttons were not being ignored, they did not exist on screen. Phones are 68% of this page traffic.

Changed The cards now open by default. A visitor can still collapse one.

Detected This page reported today as zero spend all day while the campaign was delivering. The snapshot job was already running hourly, so cadence was never the cause: it asked Meta for date_preset last_30d, and Meta's last_Nd presets exclude the current day. No schedule could have fixed it.

Changed An explicit date range ending today. Verified against the live campaign: the preset returned 09-01 only, the range returned 09-01 and 09-02. Today's spend and impressions appear on this page from now on, so a same-day reader is no longer looking at a zero that reads like a result.

Detected The page published rates but never the number the campaign is actually judged on: what a signup costs against what we will pay for one. Without it a reader could not tell a working campaign from a failing one, and the obvious reading of zero signups was that the ads had failed.

Changed The goal is now stated at the top and worked backwards through the funnel. It also computes how much traffic is needed before zero signups means anything: about 226 sessions for the first one at the required rate, and about 677 before a continued zero is evidence rather than a quiet spell. At 31 sessions the honest verdict is that it is too early, and the page says that instead of implying failure. (Thresholds as of 2026-09-02, at $0.33 a session over 31 sessions; they move with the cost per session and the live goal block supersedes them.)

Detected The landing page had exactly one conversion, the signup, and it is the slowest thing on the funnel to fire. At the required rate and this budget, the first one was not expected for about two weeks, so there was nothing to read in between.

Changed A Book 15 minutes link now sits beside Start free on every call to action, and clicking it emits a custom event through the pixel's public API. It is counted as a branch off the funnel and labelled a click, not a booked call: the calendar is a Google link and Google reports nothing back to us. The cost goal is still judged on signups.

Detected The brief was agencies with at least two people. No ad platform sells headcount targeting, so it has to be inferred.

Changed Required a second signal: tools a staffed agency buys and a solo freelancer usually does not. Reach fell from 383,000 to 135,000, so it is filtering rather than rubber-stamping.

Detected The previous agency ad pointed at a comparison page: 29.7 KB compressed against our own 25 KB budget, with a headline that promised something the ad had not.

Changed A purpose-built page at 21.6 KB whose headline is the ad headline.

Detected Clicking a question wrote it into a box and stopped, then asked for a second click to see the answer. That is a dead click, the exact thing this product detects.

Changed Selecting a question now moves the question and the answer together.

Detected On a 375px phone the only call to action sat below the answer panel, off the first screen.

Changed A button now sits under the headline, measured at 450px on an 812px screen.

Detected Our own layout check flagged logo links at 34px and a button at 26px, under the minimum Apple and Google both publish and a defect this product reports to customers.

Changed All raised past 44px, verified at five screen widths.

Detected No signup record carried a campaign tag. The capture ran only inside the app, so anyone who arrived on a marketing page and clicked through to the signup form lost it on the way.

Changed First touch is now recorded on every marketing page. There is no backfill: what was never captured cannot be recovered.

Detected Creating the ad made Meta fetch the page: 8 sessions carrying the campaign tag and a click ID before a penny was spent. Our own bot check caught 1 of the 8.

Changed Days with no spend cannot report traffic, and flagged sessions are dropped.

Detected Two Dallas trades concepts ran at 1.22% and 0.91% click-through, under the 1.5% floor, while 85% and 100% of their clicks survived the page load. The hook was failing, not the page. The agency concept then ran at 1.86%.

Changed Different audience and geography, so not a clean comparison, but enough to keep the agency angle and stop rewriting landing pages.

Detected Interest targeting alone reaches everyone who works in advertising, most of whom cannot buy software.

Changed Added a business-owner signal as a required second condition rather than a broader interest list.

Detected Our first campaign spent about $7 for zero signups: 174 clicks, 98.7% of them bouncing at under 10% scroll.

Changed Every campaign now optimises for landing page views, which only count when the page actually loads in a browser.

Detected Instagram sessions reported page-load times of 93 to 111ms on Pakistani mobile networks, which is physically impossible.

Changed Facebook feed only, on every campaign since.