Google's March 2026 Core Update: What We Got Wrong And What The Data Actually Says

14 October 2026 Last checked 2026-10-14

This page once said Google's March 2026 update landed on 19 March and ran on Gemini 4.0. Both were false. Here is the corrected version, with every figure sourced.

On 20 March 2026 this page claimed that Google's March core update had landed the day before, that it ran on a model called Gemini 4.0, and that original content was getting a 22 percent visibility boost. The update had not started. There is no Gemini 4.0. The 22 percent traces to one marketing agency's scoring rubric, not to Google.

We are rewriting the post rather than deleting it. A Gibraltar accommodation business lecturing other businesses about content honesty while leaving three fabricated claims live would be worse than the original mistake. Here is what the record actually shows, with every figure attributed.

The real timeline was 27 March to 8 April 2026

Google's March 2026 core update began rolling out on 27 March 2026 at 02:00 Pacific Time. Google's Search Status Dashboard confirmed the rollout complete as of 8 April 2026, in a note posted at 06:12 Pacific. That is about 12 days, inside the two week window Google had estimated. Search Engine Journal reported the confirmation and Search Engine Roundtable logged the same dates.

The original post went up on 20 March. It described an update that had not begun as something already reshaping results. Nothing in it could have been based on observed data, because there was no data yet. That is the whole error in one sentence.

Google published no companion blog post for this update and named no targets. Its only description was the standard one: a regular update designed to better surface relevant, satisfying content from all types of sites. Everything beyond that sentence, in our March post and in most of the coverage, was third party interpretation.

Early reporting put volatility peaks for the update at 8.7 and 9.5 on one widely watched tracker, varying by source. We are repeating that with the caveat attached rather than quoting the higher number on its own, because quoting the higher number is how the March post happened.

There is no Gemini 4.0, and Google never announced an information gain score

The March post said Information Gain Scoring was powered by Gemini 4.0. Gemini 4 had not been released in March 2026 and still had not been released in October 2026. Google's Gemini release history runs Gemini 3 in November 2025, 3.1 Pro in February 2026, then 3.5 and 3.6 Flash tier releases through the middle of 2026. Gemini 4 is in development with no announced launch date.

Information gain is a real idea with a real paper trail. Google holds a patent family titled Contextual estimation of link information gain, filed on 18 October 2018 and granted to Google LLC on 7 June 2022 as US 11354342 B2, which describes scoring a document by the additional information it contains beyond documents the reader has already been shown, and then using that score to rank or demote results. Separately, Google's guidance on creating helpful content has asked since 2022 whether a page provides original information, reporting or analysis. Those two are the closest thing to a primary source here, and a patent is not a shipped ranking signal. Google files plenty it never uses.

What does not exist is a confirmed Google ranking system by that name, switched on in March 2026, with a published effect size. Rewarding originality is consistent with what Google says it wants. The named mechanism, the model behind it and the 22 percent are not verifiable, and we should not have printed them as fact.

What the AI Overview citation numbers actually show

The one figure in the March post that traces to real research is the 38 percent. Ahrefs analysed 863,000 keywords and roughly 4 million AI Overview URLs in March 2026 and found about 38 percent of cited pages also ranked in the top ten for the same query. The remainder split almost evenly: 31.2 percent from positions 11 to 100, and 31.0 percent from beyond position 100.

That is a useful finding for a small site. Nearly two thirds of the AI Overview citations in that dataset went to pages that were not on page one. A page one position is not the entry ticket it was.

What we got wrong was the framing. We called it a flip from 76 percent, as though one study had measured both ends. It had not. The 76 percent came from a separate analysis of 1.9 million citations that looked only at the three most visible citations in each answer, which is a different question. BrightEdge, on its own dataset, puts the top ten overlap nearer 17 percent. A June 2026 CiteLens study measured about 93 percent for AI Mode, a different surface, against roughly 40 percent for AI Overviews.

Published estimates of that overlap range from 12 to 93 percent depending on surface, dataset and method. Anyone quoting a single number as the number, us included, is overstating what is known.

For a blog this size, 154 posts, the practical reading is narrower and more useful. Depth and specificity are what get a page quoted when it is not ranking first. An 874 word post with no sources, which is exactly what this page was until today, has no route into those citations at all.

We cannot tell you what this update did to our own search traffic, and that is worth saying plainly rather than filling with a guess. The Search Console service account for this site is still not connected, and the analytics key the reporting routine reads is unset, so the numbers that would answer the question are not available to us today. Anyone who publishes a confident before and after chart for a core update without saying which property and which window it covers is doing what our March post did.

How to check a claim like this in ten minutes

We should have done this in March. None of it is hard.

Start with the vendor. A claim about a Google ranking system belongs on Google's Search Status Dashboard or the Search Central blog before it belongs anywhere else. If neither mentions it, the claim is interpretation and should be labelled that way.

Check the version number. Gemini 4.0 took one search to disprove. A claim naming a model version is the easiest kind to falsify, which is why invented ones survive so long: almost nobody bothers.

Trace the number to its study. The 38 percent has a method, a sample of 863,000 keywords and a date. The 22 percent has none of those. A figure with no sample size behind it was either invented or rounded off somebody's rubric.

Check the date on the source, not just the date in the sentence. A write up dated inside the rollout window is describing partial results, and the dashboard note at the end is the only point at which the picture settles. Our March post cited no sources at all, which is its own tell: there was nothing in it to date check, by us or by a reader.

Then check whether two numbers measure the same thing. The 76 and the 38 did not. That is the error that produces the most confident wrong sentences, because both numbers are real and both are correctly quoted.

The section that broadly held up

The March post's warning about Generative Engine Optimisation shortcuts is the part research supports. Lily Ray has documented sites losing between 40 and 95 percent of traffic across the January to April 2026 window after publishing large volumes of self promotional listicles built for AI citation. Google has not confirmed a cause, so read that as correlation, not a confirmed penalty.

Hidden prompt injection is the clearest case. Microsoft's security researchers documented a technique they called AI recommendation poisoning, where instructions hidden inside a summarise button plant commands in a reader's AI assistant. Google and Microsoft have both publicly treated it as spam, and Google's spam policies cover manipulating results through deceptive means.

One claim we could not verify: several write ups say Google added a specific line about manipulating generative AI responses to those policies in May 2026. We looked and could not confirm it against the policy page. If that matters to you, read the page yourself rather than taking our word or anyone else's for it.

What a Gibraltar business should take from this

Strip out the invented mechanism and the practical advice barely moves, which is itself the point. Write what only you can write. Ahrefs' numbers say a page outside the top ten can still be cited, so specificity now does more work than domain size.

Gibraltar is a market of 37,936 residents at the 2022 census, the usually resident figure, published after early census data suggested about 38,000. Our March post said 34,000. Small market, small numbers, easy to check and easy to get wrong.

The operator's advantage is concrete rather than rhetorical. Victory Suites runs 100+ apartments here, so we know what a month in Gibraltar actually costs, which weeks fill and what goes wrong at the frontier. An agency in London can write a Gibraltar guide. It cannot write what a month in Gibraltar costs off a year of invoices.

What we changed about how this blog is written

Three things, from this post on. Every figure carries a link to where it came from, and anything we cannot source is written as unverified or cut. Dates get checked against a primary source rather than a secondary write up, because 19 March against 27 March is an eight day error that no amount of good prose repairs. And any claim naming a product version gets checked against the vendor, which would have caught Gemini 4.0 in under a minute.

This page is dated 14 October 2026 because that is when it was rewritten. The original went up on 20 March 2026. We have not quietly refreshed a date on unchanged text, which is one of the tactics the original post correctly told you to avoid.

If you run a business here and want to know why generic content stopped working, the useful reading is not an SEO blog, it is your own booking data. Ours is why pages like remote working in Gibraltar and our business and leisure accommodation guide carry figures instead of adjectives.

Victory Suites runs 100+ apartments in Gibraltar, let through 300+ booking channels, with rates quoted in GBP. If you are here for work and want a flat with a kitchen rather than a hotel room, message us on WhatsApp or book direct and we will tell you straight whether the dates you want are already busy.