Google core updates: 5 questions before you blame the update

A core update is announced, your curve drops, and the conclusion writes itself: Google has punished you. That conclusion is almost always too fast, and the cost of rushing is real. You rebuild pages that were fine, you drop a project that was starting to pay, you spend six weeks repairing something that was never broken.

To settle it with something other than intuition, we measured 16 Search Console properties across 492 days, from 21 April 2025 to 25 August 2026, and matched their daily click series against Google's official update catalogue. 9 Ranking entries fall inside that window, grouped into 8 rows because the two March 2026 updates cannot be separated, and 6 are actually measurable.

Three findings, two of which we did not expect.

• Across the 6 windows, the median share of readable properties that never move beyond their own background noise is 70.8%.
• The 41 largest breaks of those 16 months land inside a rollout window 24.4% of the time, while those windows occupy 25.2% of the calendar.
• The largest collective move of the period belongs to no update in the catalogue.

Hence the grid below. It no longer serves to measure an update's impact, as we had planned: it serves to methodically rule out everything else before blaming one.

What we measured, and how

Daily series of clicks, impressions and average position, web surface, through the Search Console searchAnalytics API. 13 properties cover the whole window, 3 have a shorter history. Zero API errors, zero properties lost, zero missing days inside the covered ranges.

For each update, we compare the median clicks of the 14 days preceding the start of the rollout against the median of the 14 days following its end. Not start against start: a rollout runs 3 to 28 days in our record, and comparing 14 days after the start would compare the before against a rollout still in progress.

Three guardrails, without which the figures mean nothing.

• The floor. Below a median of 10 clicks per day in the before window, no relative change is computed: it would be the noise of a single draw.
• The panel correction. Every gap is published twice, raw and net of the panel's median gap on the same dates.
• Own noise. The same calculation is replayed on dates with no rollout, with the same window geometry, property by property. The 90th percentile is the threshold.

A word on the denominator, because it matters. 13 properties are measurable in each window, but one of them has own noise above 100% at every rollout duration, and up to 1,028% on the 28-day windows. It is not still, it is out of reach of the measurement: we label it "unreadable" and drop it from the denominator rather than filing it with the flat ones. That leaves 12 readable properties, and every percentage below is computed on 12.

Finally, the panel is not 16 independent sites. One domain property contains 4 subdomains of the same publisher, worth 73.4% of its clicks. The reference median is therefore computed on 12 non-overlapping properties.

Finding 1: most sites do not move

Here are the six measurable windows, plus the four updates whose official dates are known but whose impact sits outside Search Console's 16-month retention.

The reading is monotonous: the bar is mostly grey, everywhere. The median share of properties that never leave their own noise is 70.8%, and 83.3% on the only two windows that straddle no known calendar effect, March 2026 and the June 2026 spam update.

Second finding, just as clear: no update pushes the panel in a single direction. Every window that moves anything produces increases AND decreases. When there is movement, it is decisive: +23.2% to +68.3% on the upside depending on the update, -16.0% to -65.6% on the downside. But it never runs one way.

Careful with one reading trap, and it is a serious one. "Up" and "down" here refer to the net gap, meaning the gap against what the rest of the panel did on the same dates. One panel property is labelled "down" on the August 2025 spam update although its raw clicks rose 10.3%: it simply rose less than the others. That is why the raw gap sits next to the net gap on every row, and why both must be read.

The corollary is that the two directions must never be averaged. On the June 2025 core update, the signed median of the four net gaps that left the noise is +2.1%, a figure that reads as almost no impact when it is the middle point between two increases of 27 and 41% and two decreases of 23 and 28%. Sizes are therefore published separately, increases on one side, decreases on the other.

Finding 2: the largest breaks do not land on updates

This one we did not expect, and it is more troubling than the first.

The script sweeps every date before looking at Google's calendar. It keeps the 3 largest breaks per property, at least 14 days apart, above the floor of 10 clicks per day: 41 breaks in total. Only then does it measure their distance to the nearest rollout.

Result: 24.4% of breaks land inside a rollout window, while those windows occupy 25.2% of the calendar. Within 7 days, 46.3% against 47.4%. Within 14 days, 63.4% against 66.5%. In other words: the largest drops of those 16 months land on an update roughly as often as chance would put them there.

The obvious objection is that the panel correction could have erased a common signal. A control ranking, computed on the raw gap with no correction at all, gives 29.3%, 51.2% and 73.2% against the same calendar shares: a tiny excess, indistinguishable from sampling noise across 41 breaks. No statistical test is computed here, and the excess is far too small for a test to change anything.

Finding 3: the largest move of the period is in no catalogue

The clearest collective move of those 16 months is not a click variation. It is a shift in the average position reported by Search Console.

Median of the reference panel: 18.67 over May to August 2025, 13.17 over October 2025 to January 2026, a 5.5-place difference between the two periods. 12 properties out of 12 move the same way, with a median shift per property of 6.16 places, concentrated on September 2025. No entry in Google's Ranking catalogue covers that move.

Its cause is undetermined, and we are not going to guess it: this data cannot separate a genuine re-ranking from a change in what Search Console counts as an impression. What is certain is the practical consequence, and it bites right now: any average-position comparison that straddles September 2025 compares two different things. That applies to any year-on-year comparison made in 2026: a 2026 SEO review setting its average positions against 2025 straddles that shift, and so does a dashboard showing twelve months of position progress. Search Console's average position is a fragile indicator for other reasons too, which we detail in our piece on the limits of rank tracking.

The 5-question grid

Those three findings set the order of operations. The grid does not measure an update's impact: it rules out the other causes, one at a time, until only the update is left, or nothing is.

In plain text, in the order they must be asked:

1. Does the break actually fall inside the rollout window, start to finish? The announcement date is not the effect date. A rollout runs 3 to 28 days in our record, and the source is the Ranking product history on the Search Status Dashboard, not a forum thread.
2. Did everyone else move on the same dates? Four of our six windows carry a clear or strong calendar drift: -13.7% over the second half of July, +55.5% in the back-to-work season, -32.3% over New Year week, +14.7% after the May bank holidays.
3. Does the gap exceed your own background noise? Replay the calculation on dates with no update, using the same window geometry. On one panel property the threshold is 12.7% for a 3-day rollout and 24.4% for a 28-day one: without that matching, you would be lenient on short updates and severe on long ones.
4. Did your positions move, or did your impressions disappear? This is the most useful discriminator in the grid, and the fastest to read.
5. What did you change, in the 60 days around the break? Redesign, migration, redirects, robots.txt, a campaign ending, a CMS change. None of that shows up in Search Console, and that is often where the answer is.

Question 4 alone is worth the detour

The largest decline in the whole portfolio inside an update window deserves telling, because it is exemplary.

One property lost 91.9% of its clicks over the May 2026 core update window. Spectacular, dated, landing squarely on a core update: case closed, apparently. Except its impressions fell 87.9% while its average position improved by 2.58 places. The blind sweep dates the break to 27 May: median clicks fall from 81.5 to 11.5 in fourteen days, while the median average position moves from 14.51 to 11.93.

Position flat or better with impressions collapsing is not a demotion. It is the signature of pages dropping out of the index. The work required is not editorial, it is technical, and it starts with URL inspection: we list the possible causes in our piece on pages Google will not index. The portfolio's biggest apparent disaster on an update date was not an update problem.

And if all five questions do point at the update?

It happens: across our six windows, between 1 and 8 properties out of 12 leave their noise each time. In that case, three things.

A core update is not a penalty. There is no infringement to correct, no reconsideration request to file. It is a re-evaluation, and the vocabulary of punishment is the surest way into the wrong repairs. Nor should it be confused with manual actions, or with questions about generated content, which we cover in our piece on AI content and Google.

Go down to page and query level. Our measurement sits at site level, and that is its most awkward limit: a site can have half its pages collapse and the other half grow without the total moving an inch. Search Console gives you both dimensions, use them before concluding anything about the site as a whole.

Do not react during the rollout. Across our six windows, a rollout lasts 17 days at the median. Changing a site while an update is rolling out makes the two effects permanently inseparable. Wait for the announced end, then the following 14 days, and measure then. What you should be watching at that point are the indicators you actually decide on, not the daily traffic curve. On substance, the signals that hold from one update to the next are demonstrated experience and expertise, which we tried to make measurable rather than declarative.

What this measurement does not say

The list is long, and it is part of the result.

• No causality. A break that falls inside a rollout window coincides with it, nothing more. A site change, business seasonality, a redesign or a technical incident produce the same signature.
• No statistical test, no p-value. The thresholds are descriptive flags on each property's own noise, and that noise is measured on contiguous, therefore autocorrelated, windows. A subsample spaced 14 days apart gives close values, which is reassuring without turning the flagging into a test.
• A narrow, skewed panel. 12 readable properties, overwhelmingly French-speaking B2B. No media site, no consumer e-commerce, no high-volume editorial site. So-called "helpful content" updates target that kind of site first: this panel says nothing about their effect on them.
• The February 2026 Discover update is not measurable here. Discover is not part of the web surface, and pulled separately the best-endowed property logs 594 Discover impressions across 16 months. No effect is inferred from it, neither positive nor null.
• The August 2026 spam update has no figure. Rollout ended 21 August, last consolidated date 25 August: the 14-day window after does not exist yet.
• No year-on-year comparison. Search Console retention is 16 months: there is no December 2024 in this data to correct year-end seasonality.
• Nothing about what changed on the sites. That is the limit that makes question 5 necessary.

One last admission, at our own expense. vydera.com is in the panel, with a median of 6 clicks per day. Of the 6 windows, 3 predate the start of its history, 2 find it below the 10-click floor, and the only one where it is measurable labels it "up" on a move from 10 to 12.5 daily clicks at the median. That is what a site too small for the question to be asked looks like. If that is you, the update is not your problem.

Running the measurement yourself

All of it is reproducible without a paid tool. The Search Console searchAnalytics API returns daily series of clicks, impressions and position over 16 months, property by property. Official dates are on Google's Search Status Dashboard, Ranking product history, and the incidents.json feed of the same service carries the end dates.

The recipe fits in five lines: pull the daily series, take the median of the 14 days before the start and of the 14 days after the end, compute the same gap on dates with no update to get your threshold, subtract what the rest of your portfolio did on the same dates, and publish the raw gap next to the net one. Budget half a day for a portfolio of fifteen sites.

The real gain is not the table. It is being able to answer, next time an update is announced, with something other than "we are monitoring it".