# What's new in seo·matrix, autumn 2026: a bigger knowledge base and data that admits what it didn't check

> October–November 2026 in seo·matrix: the SEO2BRAIN base grew to 342 pages, AI-citation tracking across several engines learned to tell a model swap from a lost citation, Bing became a second full data source, and audits say 'not checked' instead of 'clean'. What changed, and what is still unverified.

_Source: https://seomatrix.ai/blog/whats-new-autumn-2026/ · Updated: 2026-11-15_

---

Straight to the spoiler: this autumn the SEO2BRAIN knowledge base grew to 342 pages, seo·matrix learned to track AI citations across several engines without mistaking a model swap for a lost citation, Bing became a second full data source, and dozens of audits stopped printing "clean" when they hadn't actually looked. Below — what changed for a site owner, plus an honest list of what isn't verified live yet.

## How it started: a green "clean" next to a column of crosses

On 17 October I was running our audits over a read-only snapshot of a big public online store, just to compare with someone else's free AI-readiness checker. The `/seo-matrix:audit-all` summary drew a calm "✅ clean" for two audits. I opened the reports: 0 of 8 pages ready, the whole "ready" column in crosses!

How long had it been like that? The summary took its verdict from the exit code and from lines starting with a warning marker, while those two scorecards wrote theirs as a ✗ in a table and exited with code 0. Earlier that day a guard that runs every audit on an empty project had found 28 audits saying "clean" after checking exactly nothing. A zero, it seems, looks the same whether the site is perfect or the audit is blind.

That became the thread of the autumn. There are new data sources and commands, but I think most of the real work was teaching the plugin to say "not checked" out loud. Twenty-two releases, v1.226 to v1.244, between 16 October and 4 November, and five new commands. Here goes!

## 1. Knowledge: SEO2BRAIN at 342 pages

SEO2BRAIN is the knowledge base the pipeline's analysis agents read for theory. Today it holds **342 pages: 172 concepts, 98 sources taken apart one by one, and over 5,700 links between them**. New this autumn: RESONEO research on how ChatGPT retrieves pages, how AI Overviews and AI Mode choose sources, how Gemini cites, what twiddlers do and what the NavBoost and Glue schemas look like; Google's official guide to optimizing for generative AI search (May 2026); materials from *United States v. Google*; and two master prompts for upgrading a site — one aimed at Google, one at AI answers.

Each big topic got its own research post, out today alongside this one:

- [what ChatGPT actually sees](/blog/what-chatgpt-sees/) when it picks pages for an answer;
- [Google's own guide to AI search, 2026](/blog/google-ai-search-guide-2026/);
- [how AI Overviews choose their sources](/blog/how-ai-overviews-choose-sources/);
- [twiddlers, freshness and dates](/blog/twiddlers-freshness-dates/);
- [the Google leak: NavBoost and Glue, 2026](/blog/google-leak-navboost-glue-2026/).

One rule from the base is baked into the plugin: a field in a leaked schema or a line in a patent is not proof of how ranking works. Our agents may use such material, but never present it — or our own composite scores — as a confirmed Google factor.

## 2. AI citations: several engines, and a series that doesn't lie

`/seo-matrix:ai-citation` asks an AI engine your tracked keywords (or prompts by persona, buying stage and region) and parses which sources it cited: Perplexity and, through DataForSEO, Google AI Overviews, AI Mode, ChatGPT and Gemini.

- **The default engine follows the market.** For US projects it's now Google AI Mode — in the Comscore panel we used (Q2 2026, US desktop) Perplexity's share of AI prompts was close to zero. Elsewhere Perplexity stays the default. Claude is **not** switched on: we couldn't confirm its response format and price live, and guessing a format is exactly the mistake we keep writing guards against.
- **Evidence labels.** Every answer is Cited, Mentioned, Competitor cited, Not observed or Invalid; with `--repeat N` the report keeps each probe's label and says "Cited in k of N probes".
- **A model swap is not a lost citation.** The history stores the model the provider actually returned. If it changed, the report says "comparison invalid (model X → Y)" and losses become "not a finding". If half the probes came back without sources, the run is "PROBE INVALID" and never becomes a baseline.
- **Other non-losses.** An answer in the wrong language is Invalid; pages you excluded from AI features in Search Console are "excluded by owner"; links to Google itself in AI Overview sources aren't a competitor. A site on workers.dev or vercel.app is now labelled by its own name, not the platform's — before, an answer about "remote workers" could count as a brand mention (I laughed, then fixed it).
- **The engine's own searches are kept.** The queries ChatGPT and Gemini ran before answering used to be thrown away after we'd paid for them. Now they're used as facets of an existing page, not briefs for new ones — Google treats separate content for every such query, made primarily to manipulate rankings or AI answers, as scaled content abuse.

On top of the same history there's an opt-in weekly slice in `/seo-matrix:maintain` (our estimate: about $0.65 per slice for 16 prompts × 3 repeats × 3 engines). It brings events like "competitor cited for the first time", computed only against the previous *comparable* run, a share-of-voice landscape, per-prompt trends and an "AI visibility" section in `/seo-matrix:prioritize` — editorial, never an auto-action, ordered by our heuristic rather than any Google factor. And `/seo-matrix:quality-momentum --calibration` puts our own heuristic verdicts next to real outcomes; with fewer than 10 comparable pages it says "Insufficient outcomes for calibration". Will our heuristics survive that? Ask me in a few months.

Just want to know whether the engines cite you, without the plugin? The free [AI visibility check](/ai-visibility/) does the quick version.

## 3. Someone else's site in one command

Until October almost every audit needed our own articles. Now, if you need to look at a prospect's or a competitor's site, there are three $0 commands.

**`/seo-matrix:ai-ready-check <domain>`** — AI readiness without a project, at most 10 HTTP requests, each listed in the report. Every row is pass, finding, advisory or NOT CHECKED with a reason, and there's deliberately **no overall score**: one number hides which check failed, and two very different sites can both get 92 (subjective, but I've watched it happen). robots.txt is read by real tokens and roles, including control tokens that aren't crawlers at all (Google-Extended, Applebot-Extended). The report says "Checked K of 8 agent UAs"; `--deep` checks all eight, with the cap raised to 15 requests. A 429 or a Cloudflare challenge stops the run, and what's left becomes NOT CHECKED rather than "everything allowed". It also catches an llms.txt that sends agents to pages robots.txt closes — 4 links of 5 on that store!

**`/seo-matrix:snapshot-site`** takes a polite read-only snapshot: robots.txt first, Crawl-delay honoured, a stop without retry on 429, 503 or a challenge, with a list of what wasn't taken (the store said 429 on the eighth URL, so the list was long). It refuses to touch our own projects, and no command can publish or rewrite a snapshot.

**`/seo-matrix:site-audit <domain>`** runs both plus five audits of the snapshot and writes one report: a verdict, a denominator and the unchecked items with reasons for every step. Found nothing but didn't check everything — `partial`; `clean` only when there was really nothing to look at.

Along the way the robots.txt parser learned something none of our tools had noticed: on that store custom rules were glued to the Shopify template without a line break, so a crawler never sees the intended directive. Such lines are now named by line number. For a quick look at which AI bots your robots.txt lets in, there's the free [AI crawler access checker](/tools/ai-access/).

## 4. Bing: from three API methods to a second full source

If your site gets Bing traffic, this is probably the update that matters most. The Bing integration used to read 3 of 62 Webmaster API methods, so Bing data came without pages and the Bing report honestly printed "content gaps — not checked". Release v1.244 (4 November) fixed that, with response shapes captured from a live account with 19 sites:

- **Query × page.** Pages are ranked by impressions and the queries of the top 50 are read, so content gaps and page movement are judged as for Google, with a "read N of M pages" line. On one live site: 50 of 55 pages, 535 rows — and, surprise, one query shared by the English and Arabic versions of the same page!
- **Granularity checked:** query and page rows are weekly, about three months deep; daily data exists only for overall traffic.
- **What Bing itself knows:** inbound links and URLs with crawl issues, now in the dashboard's Bing card. Honest footnote: both came back empty for all 19 sites, so the filled-in shape is from Microsoft's reference, not our observation.
- **Bing keyword frequency**, $0 — Bing impressions, not Google volume, and labelled that way.

Still missing? The **AI Performance report (Copilot citations) isn't in the API** — not among the 62 methods, not in Bing's announcements. The route is the export, which `/seo-matrix:ai-visibility --import-bing-ai` reads. More in [Bing AI Performance](/blog/bing-ai-performance/).

## 5. "Not checked" instead of "clean"

Back to the opening scene — here's the fix:

- **Zero is no longer clean.** The 28 audits that called an empty project clean now print "NOT CHECKED: … nothing checked (not a clean result)" with their own exit code. Another 17 do the same when 0 of N articles are eligible, and scorecards full of crosses show as 🔎 advisory with an "N of M not ready" count. Guards fail the build if a new audit repeats any of this.
- **Languages.** A detector that judges by word lists measured nothing on a language without a list and printed that as a score. If your site is in Kazakh or Italian, you have probably seen such a score. Now it says "NOT CHECKED: language xx not supported". Kazakh joined the definition check. The Russian "это" counts as a definition only in explicit forms (ru and kk versions of the same pages now agree in 141 of 154 pairs, up from 109), and the English "X is" only when X is in the title or H1: on 68 English pages definitions fell from 47 to 15, and in 10 random changed cases the old match was never a definition.
- **The Search Console loop sees the results page.** A low-CTR query with an AI Overview is no longer a "CTR opportunity" for auto-pilot to fix with a new title — a snippet rewrite won't win the click back from an AI answer. Machine-shaped queries (`site:`, pasted prompts) are out of gaps and briefs. Google ranking incidents are written as "coincided with", never "caused", and auto-pilot waits 14 days after a rollout before refreshing.
- **Reference corpora in the tests.** Frozen public pages of two of our own sites — 46 Russian and Kazakh pages and 33 posts from this blog — go through the offline audit registry before every release, compared page by page with pinned numbers. Any drift turns the tests red.
- **New: `/seo-matrix:baseline`** — a before/after snapshot of the offline audits with a diff. An audit that stopped measuring is "incomparable", never an improvement, and snapshots from different plugin versions are flagged: the shift may be the detector, not the site.

And a confession: the Web Vitals card on our dashboard **had never worked**. The query hit a part of the Cloudflare API with no real-user data, every live call failed quietly, and the mocked tests were green! It's rewritten and checked live on seomatrix.ai. Embarrassing? Very!

> **Your scores may move — and it may not be your site** — geo-audit no longer counts structured data towards answerability (Google: "Structured data isn't required for generative AI search"), so pages without schema score higher. Russian and English pages without a target keyword lose "definitions" that never were. Articles in languages a detector has no word lists for go from a low score to "not checked". A baseline before and after an update tells these shifts apart from real ones.

## 6. Publishing gates: no invented people, reviews or hidden affiliate links

What the plugin now blocks or flags before publishing. Information for site owners, not legal advice.

- **Invented experts.** Composite "experts" (a made-up doctor with credentials) are no longer the default: only real people from the project's registry, otherwise the site's editorial team as an Organization, with no Person markup or credentials. Pre-publish blocks a composite persona on any page; on YMYL an organization, role or AI as reviewer is a blocker, and the image audit treats an AI portrait on an author page as one too. A composite survives only as an opt-in "editorial persona" with a visible note that it isn't a real person, and never as a reviewer.
- **Fixing what's live.** New `/seo-matrix:fix-attribution`: preview by default, `--apply` with a "before" snapshot, `--republish` through the gate. Tested on fixtures only so far.
- **Reviews.** Review or AggregateRating markup, or a customer-review block, without a stated source is a blocker.
- **Affiliate disclosure (FTC).** Affiliate links (only strong signals: `rel="sponsored"`, known network hosts, an Amazon `tag=`) need a sentence *before* the first link saying the site earns from them; pre-publish warns when it's missing or is just a bare "affiliate link". English and Russian are checked, others are NOT CHECKED.
- **"From $X" without fees** in a title or on a commercial page is an advisory; suggested sources now prefer .gov, .edu and regulators over statistics aggregators, labelled "verify before citing".
- **pSEO without a unique element waits** — the writer doesn't get a programmatic brief with no unique asset in the URL map.
- **AI disclosure.** `/seo-matrix:ai-disclosure-compliance` follows Article 50(4) of the EU AI Act: a site as deployer answers for AI text on matters of public interest, and only without human review or editorial control — an exemption the plugin counts only with a real named reviewer and a stated editorial responsibility.

## Before → now

| Situation | Before | Now |
|---|---|---|
| Audit on an empty project | "✅ clean" | "NOT CHECKED: nothing checked" |
| Scorecard with 0 of 8 pages ready | "✅ clean" in the summary | 🔎 advisory, "N of M not ready" |
| Article in an unsupported language | A low score | "NOT CHECKED: language xx" |
| AI engine swapped its model | "Citation lost" alert | "Comparison invalid", not a finding |
| Pages excluded from AI features in Search Console | Counted as a loss | "Excluded by owner" |
| Low-CTR query with an AI Overview | Title rewrite by auto-pilot | Separate section, auto-pilot declines |
| Bing data | Queries without pages | Query × page for the top 50 pages |
| Checking a competitor's site | Four commands, a hand-made table | `/seo-matrix:site-audit`, one report |
| Composite "expert" with credentials | Default mode | Blocked; labelled editorial persona only |
| Affiliate link without disclosure | Shipped silently | Pre-publish warning |

> **Written, but not yet verified live** — Field Core Web Vitals in `/seo-matrix:cwv-check` — the PageSpeed response shape isn't captured live yet. Per-query Yandex data — built from documentation. Bing links and crawl issues — empty for every site we have. Claude as an AI-citation engine — off. `/seo-matrix:fix-attribution` — fixtures only. One-command setup of a new EmDash site — never run live. Treat these as provisional.

## What to do after updating

1. **Take a baseline** with `/seo-matrix:baseline` and diff it after the next update.
2. **US project?** `/seo-matrix:ai-citation` now defaults to AI Mode and starts a new series, not compared with old Perplexity rows.
3. **Record your Search Console AI-features setting** — `/seo-matrix:doctor` asks.
4. **Regional domains or a non-Latin brand?** Add them to the brand settings (`brand.domains`, `brand.tokens`).
5. **Affiliate site or pSEO plan?** Mark the project as affiliate and put the disclosure before the first link; fill the "Unique asset" column, or the plan waits.
6. **Connect Bing Webmaster** and import the AI Performance export.
7. **Run `/seo-matrix:site-audit` on one competitor** and read the "not checked, and why" part first.

## So what changed, really?

Fewer green checkmarks, more "not checked" — and probably that's the right trade. A plugin that admits what it didn't measure is less pleasant to look at, but you can make decisions on it. Next update I'd like to move a few lines from the yellow box into the table. That's the story.

## Sources

- seo·matrix release notes, v1.226.0 – v1.244.0 (16 October – 4 November 2026)
- SEO2BRAIN knowledge base, state as of 15 November 2026
- Google Search Central, "Optimizing your website for generative AI features on Google Search", May 2026 — [developers.google.com/search/docs/fundamentals/ai-optimization-guide](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide)
- Microsoft Bing Webmaster API; response shapes captured 4 November 2026
- Comscore, AI prompt volume panel, Q2 2026, US desktop
- EU AI Act, Article 50; FTC guidance on endorsements and affiliate disclosure; Google Search Status Dashboard

## FAQ

### What is new in seo·matrix in autumn 2026?

Releases v1.226–v1.244 (16 October – 4 November 2026) added five commands — site-audit, ai-ready-check, snapshot-site, baseline and fix-attribution — AI-citation tracking that does not mistake a model swap or an invalid answer for a lost citation, Bing query × page data, publishing gates against invented experts, unsourced reviews and undisclosed affiliate links, and audits that print 'not checked' instead of a false 'clean'. The SEO2BRAIN knowledge base grew to 342 pages.

### Which AI engines can seo·matrix check for citations?

Perplexity, and through DataForSEO Google AI Overviews, AI Mode, ChatGPT and Gemini. For US projects the default engine is now Google AI Mode; outside the US it is Perplexity. Claude as an engine is not enabled yet, because its response format and price could not be confirmed live.

### Can I audit a competitor's site without creating a project?

Yes. /seo-matrix:ai-ready-check checks AI readiness from a domain with at most 10 HTTP requests by default and no overall score, and /seo-matrix:site-audit runs it together with a polite read-only snapshot and five audits of that snapshot, writing one report with what was checked and what was not. Both are $0; a snapshot can never be published or rewritten.

### Does seo·matrix pull Bing's AI Performance report through the API?

No. The Copilot citation report is not among the 62 Bing Webmaster API methods and is not in Bing's announcements, so the only route is the export, which /seo-matrix:ai-visibility --import-bing-ai reads. Other Bing data — queries by page, inbound links, crawl issues and keyword frequency — now comes through the API.

### Why did my audit scores change after the update?

Three changes shift numbers on their own: structured data no longer counts towards answerability in geo-audit, Russian and English definitions are counted more strictly when there is no target keyword, and articles in languages a detector has no word lists for are now 'not checked' instead of scored low. A /seo-matrix:baseline snapshot flags comparisons across plugin versions so the detector is not mistaken for the site.

