Month Four: 11 Clicks, and a Collapse We Didn't Cause

On July 27 this site's AI citations fell from 50 to 2 in a single day. I spent the following week rewriting the sourcing on all 129 posts, so I had a culprit ready. Then I checked three other domains.

TL;DR: Google impressions grew 30% to 11,595 and clicks went from 5 to 11 at an average position of 59.7, the best month this site has had. Bing AI citations totalled 1,464, up 71%, then fell 78% for eleven days starting July 27 before recovering. The same collapse hit three unrelated domains we monitor on the same schedule, one of which was not edited at all, so the cause was not on our side. Bing organic sent 17 clicks to Google's 11 on 12% of the impressions. Three of month three's five commitments did not ship, including the lead instrumentation, which means this report cannot answer the test month three set for it.

This is the fourth post in the ClickWerxs SEO build in public series. Real numbers, honest assessments, decisions reported against what the data actually did. Month one is here. Month two is here. Month three is here.

Reporting window: July 10 to August 9, 2026. Every figure comes from Google Search Console, Bing Webmaster Tools and GA4 exports dated August 15, 2026, pulled six days after the window closed so Search Console's backfill had settled. Month three's published Google figures reproduce exactly from this export, which is the check that the windows are comparable.

What month three committed to, and what shipped

Month three closed with five decisions. Two shipped.

Kill the head-term service pages as an impression source. SHIPPED. No new content was pointed at crypto, POS or interchange terms. Those three clusters fell from 68% of query impressions to 59.8%. The honest caveat is in the Google section: their share dropped because everything else grew, not because they shrank.

Rewrite the four buried pages for long-tail intent. NOT SHIPPED. No rewrite happened. The positions say so: /payments/crypto moved 79.0 to 76.7, /payments/pos-solutions 73.1 to 73.0, /payments/interchange-rates 79.3 to 79.5. That is noise, not progress.

Follow the leads, not the impressions. SHIPPED. Eleven of the 26 posts published this month went to trades and construction: concrete cutting software, roof coating software, drawing review software for home builders, builder software that doesn't fit your process, JobNimbus alternatives and the AI-for-construction cluster. That is 42% of output aimed at the vertical that produced month three's largest lead.

Ship the Stripe post month two promised. NOT SHIPPED. Third consecutive month. Month three said skipping it twice "would be a pattern rather than a tradeoff." It is now a pattern. Meanwhile /alternatives/stripe improved from position 52.7 to 49.6 with no content pointed at it, for the second month running, and still has zero clicks.

Instrument the lead path properly. NOT SHIPPED. This is the one that costs this report its conclusion. A "how did you hear about us" field exists in the site's form component previews and was never added to the live quote form or the contact form. Neither captures a landing page or a referrer. Month three called untraceable leads "a reporting failure, not just an attribution limitation," committed to fixing it, and did not.

What we published in month four

Twenty-six posts between July 10 and August 9, a pace of 5.9 per week. Month three ran 4.2, month two ran 5.0.

Those counts come from git, not from the date field on each post, and the distinction matters this month. On July 29 the custom app cluster and several construction posts were backdated onto free date slots, so the site now displays 23 posts inside this window while git records 26. The date field records when a post is presented as published. Git records when it went live. Month three's published figure of 18 posts reproduces from git and no longer reproduces from the date field, which is exactly the kind of drift that makes a monthly number quietly stop meaning what it meant last month.

Composition: 11 trades and construction, 5 custom apps, 4 AI SEO, 3 payments, 2 AI benchmarks, and month three's own report.

What the Google data shows

Month four (July 10 to August 9): 11,595 impressions, 11 clicks, 0.095% CTR, average position 59.7.

Month three (June 10 to July 9): 8,950 impressions, 5 clicks, 0.056% CTR, average position 63.6.

Impressions grew 30%. Clicks went from 5 to 11. Position improved 3.9 places. On every metric this is the best month the domain has recorded.

It is also six additional clicks. A 120% increase on a base of five is a sentence that flatters the number rather than describing it, and the rest of this section is why I am not leading with it.

Across the trailing 90 days, May 14 to August 13, the site recorded 27 clicks on 31,133 site-level impressions. The Pages report shows 32,761 page impressions across 190 pages for the same window, because the site chart counts one impression per query while the Pages report counts one per page shown. The percentages below use the page-level total, which is the only correct denominator for comparing pages.

The four buried pages from month three:

Their share of page impressions fell from 60.3% to 55.6%, which reads like the month three decision worked. It did not. In absolute terms these four pages grew from 15,376 impressions to 18,203. Their share dropped because the rest of the site grew faster. Declining to feed a page does not make it stop collecting impressions, and I reported that share as a win in my own notes before checking the absolute number.

The genuinely good news is underneath. The English blog went from 6.5% of page impressions to 13.6%, from roughly 1,657 impressions to 4,443, and from 2 clicks to 5. The homepage holds 348 impressions and 8 clicks at position 8.57, up from 265 and 7 at 8.97.

Now the number that governs everything else. Across the entire 90-day query export, 44 queries sit at position 20 or better. They account for 333 impressions, 1.3% of all query impressions, and produced zero clicks.

That is the whole problem stated in one line. Ninety-nine percent of what Google shows this site is at a position where clicks do not happen, and the one percent that is well positioned is well positioned on queries nobody searches: "melio level 2 data level 3 data" at position 3.1 on 13 impressions, "square issuing v2" variants at positions 5.7 and 5.9, "zoho command center" at 9.4 on 23 impressions, "jobnimbus alternative" at 18.5 on 27.

Two caveats on the query data, both of which cut against reading too much into it. The export is capped at 1,000 rows covering 24,913 of 32,761 page impressions, and it shows 2 of the 27 clicks, because Search Console anonymizes low-volume queries. And no branded query appears anywhere in it, while the homepage collects 348 impressions at position 8.57 with a 2.3% CTR, which is what branded search looks like from the page side. Both facts are true and only one of them is visible in the query report.

One finding I did not expect: five of the site's 27 clicks came from Spanish pages, at positions 7.0, 9.8, 11.4 and 16.4. SPEI, CoDi and DiMo for Mexican merchants alone produced two clicks on 97 impressions. The Spanish blog is 36 posts against the English site's 101, roughly a third of the output, and it is ranking where clicks are possible, which no English cluster except the homepage currently does.

The Bing collapse

Month four: 1,464 citations, 47.2 per day, 9.2 cited pages per day.

Month three: 855 citations, 28.5 per day, 7.1 cited pages per day.

Total citations rose 71%, the daily rate rose 66% against month three's 30 days, and breadth rose 28%. Those averages hide the only thing that mattered this month.

This was not a slide. On July 26 the site logged 50 citations across 12 pages. On July 27 it logged 2 citations across 1 page, and stayed flat on the floor for eleven days before returning to 70 on August 7. Citations fell 78% and breadth fell 83%, to a level below anything recorded since May.

I had an obvious culprit. Between July 29 and August 1 I rewrote the corpus: a sitewide brand spelling correction across 251 occurrences, publish dates reassigned across every post in both languages, Sources blocks added to all 129 posts, every competitor citation removed, hreflang drift and 13 broken links fixed, dateModified stamped on 31 posts, and every author linked to a single credentialed Person entity. Four days of rewriting the metadata, sourcing and schema of an entire site, landing in the middle of the worst citation window it has recorded.

The story writes itself, and it is wrong.

The chronology does not fit. The cliff is July 27. The first sweep commit is July 29 at 08:46. The only commits on July 27 were two merges at 20:49, too late in the day to explain a full day that recorded 2 citations.

Nothing was removed. The obvious mechanism for losing citations is losing URLs. The corpus went from 107 posts on July 25 to 129 by July 30 and never shrank. It grew through the entire collapse.

And it was not just us. The same drop and the same recovery, on the same dates, appeared across three unrelated domains we monitor: a roofing contractor, an IT services company, and an e-commerce operator. Different verticals, different sizes, no shared content. Two of the three were being edited during that window. One was not touched at all, and it collapsed on the identical schedule.

That last one is the control, and it settles the question. Whatever happened between July 27 and August 6 happened at Bing, not in my repository.

I am publishing the wrong hypothesis alongside the right one because the gap between them is the actual lesson. Watching one property, I had a tidy, confident, causally plausible story: I touched everything, everything fell over. The conclusion I would have drawn is that mass sourcing corrections cost AI citations, and the behaviour that follows is being slower to fix factual errors on a live corpus. That is a bad outcome produced by good-faith analysis of insufficient data.

Four properties is still a small sample, and Bing publishes no incident history that would let me confirm anything on their side. I cannot tell you what happened. I can tell you it was not the sweep, and that I only know that because I had somewhere else to look.

Bing organic moved the other way and did not dip. Seventeen clicks on 1,436 impressions this month, against 8 clicks on 477 in month three. Across the trailing 90 days, 30 clicks on 2,290 impressions.

Which produces the comparison that has held all four months and got wider this one. Bing sent 17 clicks. Google sent 11. Bing did it on 1,436 impressions against Google's 11,595, roughly 12% of the volume. Google shows this site to eight times more people and sends fewer of them, because the impressions it serves are on page six.

Leads, and the instrument we did not build

GA4 recorded 9 generate_lead events from 6 users across the window, alongside 9 form starts from the same 6 users, 1,035 page views, 942 sessions and 867 users, of whom 862 were first-time visitors. Lead events more than doubled on a per-day basis against month three, from 0.12 a day to 0.29.

Before anyone reads that traffic as growth, one of those numbers is not real. GA4's own anomaly detection flags August 4, when page views hit 229 against an expected 29, and attributes it to visitors from Singapore rising from 3 to 218 week over week. That is close to a fifth of the month's page views and a quarter of its users arriving in a single burst from a market this site does not serve and has no content for. The two spikes visible in the chart below are that event and a smaller one on July 22. Neither is audience.

The arithmetic is worth doing rather than waving at. Raw, the site averaged 33.4 page views a day against month three's 21.8, which reads as 53% growth. Subtract the 200 excess page views GA4 flagged on August 4 and it is 26.9 a day, or 24%. Both are real numbers. Only the second one describes the audience.

I cannot tell you how many of those were organic, which verticals they came from, or which page produced them.

That is not an attribution limitation. It is the direct consequence of not shipping the fifth commitment. Month three identified this exact gap, named it a reporting failure, committed to closing it with a "how did you hear about us" field and landing-page capture, and month four shipped 26 blog posts instead. The forms collect the same nothing they collected in July.

So month three set a test for this report: if month four produced citation growth with no movement in branded search or leads, the bet is wrong and I said I would say so.

Citations grew 71%. Branded search moved a little: homepage impressions up 31%, position from 8.97 to 8.57, clicks from 7 to 8. On leads I have no answer, and the reason I have no answer is a decision I made and did not execute. A falsification test you disable is not a test.

The counterargument, and it is stronger than last month

Month three's skeptical read was that citations are a vanity metric with no proven path to revenue. Month four handed that argument a gift: the channel fell 78% for eleven days for reasons outside my control, recovered on its own, and I cannot point to a single business consequence in either direction. A metric that can halve and restore itself without anything happening to the business is not yet a metric anyone should manage against.

Month four adds a second charge that is harder to answer.

Three of five commitments did not ship. The one that would have made this report conclusive is among them, and the one that has now slipped three months running is the single clearest gap in the data. Output went up 40% while execution against stated decisions went down. Writing 26 posts is easier than rewriting four pages and adding two form fields, and the month four allocation shows me choosing the easier work while reporting it as progress.

The defensible part: the trades cluster shipped as promised, blog impressions doubled as a share, position improved almost four places, and Google clicks moved for the first time since month one. Those are the outputs the strategy predicted, on the schedule it predicted.

What changes in month five

Two form fields, before any new post. Source dropdown and landing-page capture on the quote form and the contact form. This has been outstanding for a month and it blocks every attribution question in this series. Nothing else ships until it does.

The Stripe post, or the commitment gets retired in writing. Position 49.6 on 2,358 impressions with no dedicated content, unexploited for three months. Month five either ships it or states plainly that the series is dropping it, because carrying a commitment nobody executes costs more credibility than abandoning it.

Rewrite two of the four buried pages, not all four. Month three committed to all four and delivered none, which suggests the scope was the problem. /payments/interchange-rates at position 79.5 and /payments/pos-solutions at 73.0 get rewritten for long-tail intent. The other two wait.

Report Spanish separately. Five of 27 clicks at positions 7 to 16 came from a section this series has never broken out. It is outperforming the English site on every metric that involves a click, and I have been reporting it inside a blended total that hides it.

Filter the analytics before reporting from it. One flagged day carried 19% of this month's page views from a market the site does not serve. Month five adds a filtered GA4 view that excludes known non-target regions and bot-shaped sessions, so the reported baseline is the audience rather than the total.

Track citations across every property, not just this one. The July collapse was only diagnosable because three other domains were visible. From month five the AI Performance series for every property we monitor gets pulled on the same day into one sheet, so a platform-level event is obvious in an afternoon instead of a fortnight. A single-site dashboard cannot distinguish "I broke it" from "it broke," and that distinction changes what you do next.

No forecast for any of it. Four months of data on this domain does not support predicting a fifth, and a number I have no basis for would be the least honest thing in an otherwise honest report.

The next report goes up on September 25.

Frequently Asked Questions

Does stamping dateModified across many posts at once carry a risk?

On this evidence, no. That was my working assumption for a week, and the cross-domain data killed it: a property that received no edits at all fell on the same days by the same amount. Nothing in Google's or Microsoft's documentation warns against bulk metadata correction either. If you are sitting on factual errors in a live corpus, fix them. The thing worth changing is not your edit schedule, it is having a second property to check against before you blame your own deploy.

How do you tell a real AI citation dip from a reporting lag?

Two checks. First, whether volume and breadth move together and whether the data backfills: on this site volume fell 78% and breadth 83%, the recovery held rather than appearing retroactively in the old dates, and nothing refilled. Second, and far more decisive, compare against another property. A lag or a platform event hits several unrelated domains at once, and a problem you caused does not. That comparison took ten minutes and it was worth more than a week of reading my own commit log.

Why report git ship dates instead of published dates?

Because the published date is editable and the commit is not. Any site that backdates posts onto free slots, which is a normal editorial practice, breaks the link between the date field and when content became crawlable. If you report cadence from the date field you will eventually report a month you did not work. Pull it from version control, or from your CMS's created-at rather than its display date.

Is 11 clicks a month worth reporting at all?

At this size the click count is nearly meaningless on its own and the position distribution is the real signal. Eleven clicks tells you almost nothing. Forty-four queries at position 20 or better, out of a thousand, tells you the site is being shown constantly and ranked nowhere, which is a diagnosis you can act on. Report the number that changes your next decision.

Should a second-language version of a site get its own SEO reporting?

If it converts differently, yes, and this site is a case for splitting the reports sooner than we did. Spanish pages here are about a third of the post count and produced five of 27 clicks at positions between 7 and 16, while the English site outside the homepage produces clicks almost nowhere. Blended into one total, that difference is invisible. A language version is a different index, different competition and often a different funnel, and averaging the two hides whichever one is working.

The most useful number in this report is the one that made me look worse: three of five commitments did not ship, and the missing one is why this report has no conclusion about leads. If you want SEO reporting on your own site done at this level, including the months where the honest answer is that the work did not get done, ClickWerxs builds and manages it.

Operator opinion. Not a guarantee of results for your site. SEO timelines vary by domain age, niche, and content quality. All first-party figures come from Google Search Console, Bing Webmaster Tools, and GA4 exports dated August 15, 2026.

Kaleb Dickhaut — Founder, ClickWerxs Kaleb built ClickWerxs from the ground up, from payment processing ISO to the Command Center platform to the AI SEO methodology this blog runs on. He has onboarded hundreds of small businesses onto payment and CRM systems. linkedin.com/in/kaleb-dickhaut

Sources

1. ClickWerxs first-party data — Google Search Console, Bing Webmaster Tools and GA4 figures in this post are measured from ClickWerxs's own properties over the stated periods, from exports dated August 15, 2026. Google Search Console, Bing and GA4 figures all cover July 10 to August 9, 2026. Page-level and query-level breakdowns come from the trailing 90 days, May 14 to August 13, on the same basis month three used, so that month-over-month page comparisons are like for like. The August 4 traffic anomaly is reported from GA4's own anomaly detection. They are a past result for one domain in one vertical and are not a projection of what any other site will achieve. Operator data.
2. Publishing cadence figures are measured from this site's git history rather than from post display dates, for the reason given in the text. Operator data.
3. The cross-domain comparison in the Bing section is an operator observation across three additional properties ClickWerxs monitors, in roofing, IT services and e-commerce. They are unrelated to each other and to this site, and are described by vertical only; no client is identified. The observation is that the citation drop and recovery occurred on the same dates across all four properties, and that one of the three received no content or schema changes during the period. Operator data.

ClickWerxs sells SEO and AI visibility services and earns revenue from those engagements. First-party figures are past results for this blog and are not a promise of future performance. This is operator opinion, not professional advice.