From 4,788 to 1,545 impressions: the test that separates a real drop from a mix change

In January 2026, marketingnativo.com pulled 4,788 impressions a month. In July it pulled 1,545. The median position of its pages went from 8.2 to 24.7, and the share of impressions coming from beyond position 30 climbed from 6.3% to 26%. It is my own site, so I can show the numbers unrounded and include the part where my own tooling handed me the comfortable answer three times in a row.

Quick answer

• The most quoted metric is the least reliable. A worsening average position can mean a real loss, or it can mean new bad queries showed up that were not there before.
• The test that separates them is a fixed cohort: take the queries from the first month and look up those same queries later, not the top queries of today.
• Here the drop was real, and clicks carry the proof: 34 a month down to 7. That is the one column a Google logging error did not touch.
• Google's own guide is the best result for this question, and this article starts where that guide stops.

The bottom panel is what holds the diagnosis up. Google states its logging error affected impressions, CTR and average position, and not clicks, so the fall from 34 to 7 does not depend on it.

Why did my organic traffic drop?

Organic traffic drops for three reasons that get diagnosed differently: you lost rankings on queries you already had, those queries lost search demand, or the mix changed and your average position got worse without anything falling. The test that separates them is a fixed cohort: measure the same first-month queries in the last month.

What kind of problem this is

The dominant mental model treats average position in Search Console as a thermometer. It goes up when things go well, down when they go badly, so a move from 8 to 24 reads as an emergency.

It is not a thermometer. It is an average over a population that changes on its own. A hundred new queries at position 60 will sink it without a single page losing a spot. The reverse works too: a site that loses its best queries can watch its average improve, because what survives is the handful that always ranked well.

So this case is not about how I recovered. It is about the step before that, which almost nobody shows: how you find out what happened before touching anything.

Where the English-language answer already is good

This is worth saying plainly, because it is not what I found searching the same question in Spanish. In English the first result is Google's own guide to debugging Search traffic drops, and it is genuinely useful. It lists six causes, tells you to read the Performance report across a 16-month window, points you at Google Trends for demand shifts, and it explicitly tells you to check the Data Anomalies page. Two of its six causes are technical, and the one that hides best is covered in why is my page not indexed.

The rest of the first page adds a heuristic worth keeping: if impressions hold and clicks fall, your rankings are probably intact and something above your result is eating the click. If both fall, you have a ranking problem.

My site failed that heuristic in the least useful way. Both fell. The heuristic correctly says "ranking problem" and stops, and that is not a diagnosis you can act on: it cannot tell you whether your pages lost their spots or whether Google started showing you for junk queries while the old ones went dark.

Google's guide gets closer. It says to filter the Performance report one query at a time and find the ones that lost clicks and impressions. What it does not say is which queries to pick, and that omission is the whole ballgame.

The fixed-cohort test

Take the queries that brought the most impressions in the first month of your series, write down their positions, and look up those same queries in the last month. Not the best ones now: the ones from before.

If the drop were a mix change, the old queries would still be where they were and the average would have moved because of the new ones. If they fell too, the loss is real. It is a cohort in the epidemiological sense, and it kills the error that makes most monthly reports useless: comparing two populations and calling it a trend. Getting the raw numbers out is covered in how to set up Search Console.

What it returned here

The thirteen queries from January, measured against July, which is the last complete month:

• Six fell five positions or more. The biggest one, a definitional query about the software the blog covers, went from position 8.3 to 21.3 and from 106 impressions to 3.
• Six went dark entirely, meaning zero impressions.
• Cohort impressions went from 434 to 72, down 83%, while the site's total for named queries fell 30%.
• The average shift, weighted by the impressions each query used to bring, is 15 positions.

The cohort fell much harder than the site did. That rules out a mix change and it also rules out a general decline: the loss is concentrated in the queries that were holding the site up. The page-level report agrees. The homepage sat at position 9.6 in April and 26.1 in July.

When NOT to use this test

When the first month has very few impressions. A cohort built on thirteen queries with twenty impressions apiece measures noise.

When the site changed topic on purpose. Old queries going dark is then the intended outcome, not a symptom.

When the drop lines up with a technical change you already shipped. Fix that first and re-measure, because a broken template moves everything at once and the reading is in Core Web Vitals in Search Console.

When the last month is still running. That was my mistake and it gets its own section.

When you suspect search demand itself moved. Search Console does not carry query volume, so this test cannot rule that one out. Google Trends can, and Google's guide is right to send you there.

Three bugs in my own tool

The script that runs this test is mine and had been in use for months. Here it returned "the cohort holds", which translates to: nothing happened, it was a mix change. That contradicted the page-level report, so I opened the code.

First bug: a query that disappears did not count as a fall. The tally only looked at queries with data at both ends, so going from position 45 with 42 impressions to not existing scored zero. With six of thirteen gone dark, the verdict came out reassuring by construction.

Second bug, and I introduced it fixing the first. I counted disappearances as position losses, and the tool declared "real drop" on westchesterride.com, which over those same months went from 3,877 to 102,472 impressions. On a growing site the weak first-month queries evaporate on their own: that is churn, not loss. A false negative traded for a false positive.

The fix that worked was splitting the two questions instead of fusing them into one verdict: positions get judged using the queries still alive, weighted by what they used to bring, and churn gets reported separately with the site's own trajectory next to it. A query going quiet means opposite things depending on whether the site is growing.

Third bug, the dumbest one, and it moved the answer the most. The series ended on the month in progress, eleven days in, where positions come from queries with one or four impressions and swing on their own. Closing on July the shift is 15 positions and the verdict is a real drop; closing on eleven days of August it is 3.5 and the verdict is mild deterioration. The second was the artifact, and I nearly wrote this article around it.

The warning that almost invalidated all of this

Before publishing I opened the Search Console Data Anomalies page, where Google logs its own failures. One applies directly: a logging error kept Search Console from recording impressions accurately from May 13, 2025 to April 27, 2026, affecting impressions, CTR and average position.

My series crosses that date down the middle: January through April sits inside the window, May through July outside. Comparing January's position to July's compares two measurement regimes, which is the sin this article is about, with the official source one click away.

What saves the diagnosis is the one column Google says was untouched: clicks. January had 34 and July had 7, down 79%. That series does not depend on the error, so the drop stands even if you threw out every position figure above.

Mistakes that repeat

• Diagnosing from average position. It is the number that leads most reports and the one that survives a question the worst.
• Comparing this month's top queries against last month's. Two different populations; what comes out of that is not a trend.
• Ending the series on the current month. The last few days bring tiny samples and positions that swing.
• Trusting a tool because you wrote it. Mine handed me the answer I wanted twice, for two different reasons.
• Reaching for causes before confirming the effect. Four of the six causes in Google's own list fit my site at once, so the list cannot discriminate on its own.

• Confusing a drop with a penalty. Different things with different symptoms, and the difference decides whether you have anything to undo. Agency reporting rarely draws it, which is one of the tells in is my SEO agency working.

What happened, as far as it can be claimed

The jump is in June. Median position per page moves from 9.1 to 20.2 and impressions past position 30 go from 0.9% to 34.8%, in the same month. Google confirmed a core update that started May 21, 2026 and took nearly twelve days to roll out, which makes June the first complete month after it.

A date coincidence is a date coincidence. Google does not announce which sites an update affects, so this is consistent with it and is not evidence of it. What the cohort supports is a real loss of position, concentrated in the definitional queries of the niche, and not a mix artifact.

These numbers cannot rule out a second reading. The queries that fell are all the "what is X" shape, which is what generative summaries absorb first. Telling the two apart means checking whether those queries carry a summary, and the mechanics are in AI Overviews and local search.

Data and transparency

Figures come from Search Console, property https://marketingnativo.com/, pulled through the API on August 13, 2026 with a read-only credential. It is my own site, so it appears by name and unrounded.

Search Console numbers change when you ask again. The same closed month returns different values days apart because Google backfills, so every figure carries its consultation date.

Scale matters. 28,053 impressions over twelve months is a small site, where percentages move easily. The automatic cohort also drags in brand queries.

August is not extrapolated. From the 1st to the 11th it had 2,492 impressions, more than all of July, which points at a partial recovery. Eleven days are not a month.

The 2026 update dates and the logging anomaly both come from Google's public communications, checked on August 13, 2026.

What this changes

The useful conclusion is not that a site of mine lost positions in June. It is that I was one decision away from publishing three different diagnoses from the same data, and the difference was not in the data. It was in how I closed the window and what my script counted as a fall.

That puts the cause lists in their place, including the good one Google publishes. The problem is not that the causes are wrong: it is that arriving at causes before confirming the effect starts at the end, and a site that drops fits four or five boxes at once. Picking one becomes picking the one you already suspected.

The uncomfortable part is the other one. I wrote the tool, used it for months, and it handed me the comfortable answer the first two times I pointed it at something that stung. A check that has never failed is not validated, it is untested.

Frequently asked questions