We Reverse-Engineered Google's October Re-Rank. Here's Who It Trusts Now.

Google's October 2026 re-rank scored sites, not just pages, and the sites it now trusts on startup and venture questions are primary sources, legacy wires, Reddit, narrow data brands and Wikipedia. Identity predicted rank. Word count didn't.

I run a site that lives on startup and venture money questions. What did this company raise? What's it worth? When is the IPO? Those queries are my lab.

In mid-October the lab changed. Google re-ranked the whole category. It affected us, like it affected plenty of people in this space. So we did what any analyst should do when the market moves: we stopped guessing and pulled the data.

How we did it

We analyzed ranking data across thousands of queries in our category from Google Search Console. We ran repeated clean-profile live-SERP panels on the money queries (IPOs, valuations, funding rounds), captured on October 31 and November 9, plus daily canary tracking in between. We tore down the winning pages against the pages they displaced: word count, schema, structure, sourcing and bylines. Then we spot-checked winners against archived copies of the same pages to see what had changed. Several hadn't. The page that now leads "plaid ipo" hadn't changed in over a month. The ranker changed, not the pages.

Finally, we ran the same questions through Bing's Copilot, ChatGPT and Perplexity to see whether the machines agreed with Google. They didn't.

Finding 1: Google scored sites, not pages

The opening clue was the timing. The moves weren't gradual and they weren't page by page. Whole domains moved on the same day, across unrelated queries.

That points to something specific. Google's ranking systems guide says it plainly: "Site-wide signals and classifiers are also used." Google still ranks mostly page by page, but a site-level judgment sits on top. When that judgment changes, every page on the domain moves together.

Look at the calendar. Google's September 2026 spam update ran from September 24 to October 8. Its scaled-content policy targets sites that publish lots of pages mainly to rank. And the November 2026 core update started on November 10 and is rolling out right now, with Google saying it could take up to two weeks.

The working model: Google decided which domains it trusts on which topics, then let pages compete within that frame.

Finding 2: The new trust stack

When we sorted the winners in our panel, five types showed up again and again.

Primary sources. The company's own page. An investor-relations page or a press release now regularly beats 5,000-word analyses written about that company. Google would rather send you to the horse's mouth.

Legacy wires. The newswires and the big business desks. They've held the same identity for decades and Google treats that as collateral.

Human forums. Reddit, everywhere. In our panel it sat at #1 on several commercial queries, where it would have been a joke a few years ago. DataForSEO's 2026 data has Reddit among the top sources Google's AI Overviews cite, second only to YouTube.

Narrow data brands. This one surprised us most. Short pages from companies whose whole identity is a proprietary dataset, often a few hundred words with a chart, beat longer, better-cited essays. We found a 400-word brief with no byline outranking a 4,800-word sourced analysis of the same company.

Wikipedia. Still the default answer on any entity question.

Now the counterintuitive part. In the teardowns we checked the usual suspects: word count, number of citations, tables, freshness. None of them predicted rank. A 5,500-word sourced analysis lost to a 4,200-word lead-gen guide. Thicker was not better.

What did predict rank was identity. Who are you, and has Google already decided you're the authority on this topic? Content quality matters only inside the bracket that identity sets.

Finding 3: How the winners are built

Identity is the gate, but the winners also share a structure. This is the part you can copy.

They answer in the opening sentence. The number, the date, the name, before any context. A page that makes you scroll for the answer looks like it was written for a crawler.

One page per intent. The winners don't have three pages fighting over the same question. Near-duplicate pages split the signal and look like scale.

One consistent entity. One domain, one organization, a consistent publisher identity in the structured data, real authors with real histories. A site with no clear owner has nothing for trust to attach to.

They own their numbers. The data brands don't summarize other people's figures. They publish their own. If your number is just a copy, the engine can cite the original instead.

Stable URLs, honest dates. No slug churn, no "updated today" stamp on a page that didn't change. More on why below.

Schema only where it's true. Dataset markup on real datasets, ItemList on real lists. Markup is a claim, and false claims get noticed.

Where we beat incumbents, it came from one move: getting closest to the primary document. We read the S-1s and other filings that few people actually go through. Our piece on Wella's IPO filing ranks #1, and Google's AI Overview cites it. When competition on a filing is thin and you're nearest the source, you win the blue link and the AI citation together.

There are second-order signals here we're still validating. That one we're keeping for now.

Finding 4: How a classifier sees your site

Google publishes a list of self-check questions in its helpful-content guidance. Read them as a spec, not as advice. Each one is something a classifier can measure across a domain.

Here's the test I use. Call it the swap test. Take any paragraph and replace the company name with a competitor's. If the paragraph is still true, it's filler. A site where most pages pass the swap test is a template, and templates are what scaled-content systems are built to find.

Other signatures are easy to detect at scale. Clusters of pages with the same skeleton. Sudden volume spikes. Dates bumped in sync across hundreds of pages. Google asks outright: "Are you changing the date of pages to make them seem fresh when the content has not substantially changed?"

Why does this work at the site level? Because the classifier is estimating a prior: how likely is any page from this domain to be worth showing? One burst of templated output can drag down that estimate for your best work too.

The tell usually shows up before the re-rank, if you know where to look. I'll leave that there.

Finding 5: AEO is three games, not one

Everyone says "AEO" like it's one channel. Our cross-engine pulls say otherwise.

Start with Google's own AI layer. In our money-query panel, an AI Overview appeared on 30 of 30 queries. Industry-wide, DataForSEO's 2026 data has AI Overviews on about 61% of US results. When one appears, the #1 organic result's click share falls from roughly a third to about 21-24%. Ranking first is worth less than it used to be.

But being cited inside the overview is a separate asset. Google's Generative AI report in Search Console credited our pages with 1.75 million impressions from AI Overviews and AI Mode over one summer, growing about 5x in three months. Most of those came from explainers about how AI companies make money and from data pages, not news.

Bing's Copilot made different calls. Its citation mix favored comparison pages and fund-performance analyses, and its trust decisions didn't track Google's re-rank at all. Same web, different judgment.

ChatGPT and Perplexity made a third set of calls. They cited our data pages, the valuations and the funding numbers, and they ignored our "best X" listicles.

So there's no single AEO playbook. Three engines each decide independently what a source is. The one thing they agree on: they cite sources of facts, not SEO pages.

What to do with this

Be the citable source for something narrow. One filing type, one dataset, one market. Narrow and owned beat broad and borrowed in every engine we tested.

Structure for extraction. Answer first, numbers you own, sources linked, a date that means something. Then license it so people can reuse it. Our worked example is the open funding dataset on valueaddvc.com/statistics: verified rounds under a CC BY license. Anyone can reuse it as long as they credit us. That credit is the product.

One entity, one domain. Consolidate. Every satellite site and doorway page is one more chance for a classifier to misjudge who you are.

Treat patience as structural. Site-level judgments get recalculated on update cycles, not on your schedule. Google's core-update guidance says it "could take several months" for its systems to confirm a site has changed. Mid-rollout movement often reverses. Measure after the update finishes.

Keep a channel no algorithm owns. Video, a newsletter, a feed people chose to follow. We publish ours daily on Pulse.

For twenty years the search game rewarded publishing more.

The next one rewards being worth citing. Pick which game you're playing.