What did Google change in its helpful content guidance in October 2026?

Google rewrote its guide to creating helpful content around the main content of a page, which it defines as any part that directly helps the page achieve its purpose. That includes primary text and media, interactive tools, useful user contributions, tabbed or expanded sections, and page titles and headings. It now judges that content on four attributes: effort, originality, talent or skill, and accuracy, and says attribution or credit to other sources doesn't replace original effort. The guide says automatically generating pages from feeds, or using generative AI to produce large amounts of text without manual oversight or curation, represents little to no effort. A separate addition says fabricating creator profiles with AI-generated headshots, made-up names or false credentials is a form of deception. Google's AI content guidance was updated on October 1 to say every AI-generated element should be fact-checked before publishing, including titles, meta descriptions, structured data and image alt text.

Is the September 2026 spam update finished?

Not officially, as of October 7 2026. Google started the September 2026 spam update on September 24 and said it could take up to two weeks. Barry Schwartz tracked waves of movement on September 25 to 27 and September 30, and what looks like a final phase between October 4 and 6. Google had not posted a completion notice on its Search Status Dashboard when this was written. It is the fourth spam update of 2026.

How long does it take to recover from a Google update?

According to timing figures Gary Illyes presented at Google's Search Central Live Deep Dive in Barcelona on October 2 2026, as reported by attendees, recovery from a core update typically takes three to six months and can take six months to a year. The same figures put typical discovery of a new URL at about 20 hours and a site move at one to three months. Google has not published the slides, and Illyes described the numbers as an exercise in whether the audience could relate to internal data, so treat them as ranges rather than promises.

Does Gemini add UTM parameters to the links it sends?

Sometimes, as of October 2026. Some Gemini referral links now carry utmsource=gemini.google.com and utmmedium=ai-assistant. Google's John Mueller acknowledged the change when it surfaced on October 1, but Google has not documented when the tags are added. On the small business websites LovedByAI monitors, the first tagged Gemini visit arrived on October 4, and through October 6 only about one Gemini visit in twenty-five carried the tags. Every tagged visit also carried the gemini.google.com referrer, so analytics that already read the referrer were not missing these visits.

Can ChatGPT find my staging site or subdomains?

It can if they are public. ChatGPT increasingly searches with the site: operator, which scopes a search to one domain, and Lily Ray showed on October 6 2026 that the pages it retrieves this way include subdomains such as help centres and documentation. Anything public on a subdomain, including a staging copy of your site, can end up informing an answer about your business. Password-protect or noindex staging sites, and review help and documentation subdomains as carefully as the main site.

Can you see AI Overview and AI Mode traffic in analytics?

Not separately. A click from a Google AI Overview or from AI Mode arrives with an ordinary google.com referrer and no tag, so analytics tools count it as organic Google search. ChatGPT tags its links and Gemini began tagging some of its links in October 2026, which means any report of AI referral traffic measures only the assistants that identify themselves. That is why such reports show ChatGPT with nearly all AI traffic: Google's own AI surfaces are in the organic line, not missing.

Is OpenAI watermarking ChatGPT text?

In the EU, yes. OpenAI said on October 5 2026 that it will add an invisible watermark, called textGrain, to text generated by ChatGPT and Codex for EU users over the following weeks, to meet the EU AI Act's transparency rules. API developers anywhere can switch it on, and it is off by default. The watermark survives copy and paste but weakens with editing: OpenAI's tests found detection falling from about 92% to 66% when a tenth of the words were swapped for synonyms. The detector is limited to approved researchers and organisations, so neither Google nor a competitor can run your pages through it.

AI Search News: October 2026

The first week of October was Google telling site owners, in writing and from a stage, what it is going after next: pages produced at scale without real effort, and authors who do not exist. Underneath it, Gemini started labelling the visits it sends you, and ChatGPT's interest in your site kept getting more selective. For the longer-run picture behind these monthly readings, our 2026 AI traffic study benchmarks what AI assistants actually send to small business sites over a full window.

October 2026 at a glance

Sep 30CloudflareCloudflare opens a Pay Per Use beta that pays publishers when an AI product uses their content, not when it crawls.
Oct 1GoogleGoogle's AI content guidance now says to fact-check every AI-written title, meta description, structured data block and alt text.
Oct 1GoogleGemini starts adding UTM tags to some of the links it sends visitors through.
Oct 1OpenAIChatGPT adds virtual try-on for clothes and a place to save favourite products.
Oct 1GoogleGoogle's helpful content guidance is rewritten around main content (effort, originality, skill, accuracy) and calls fake author profiles deception.
Oct 2GoogleGary Illyes tells Search Central Live in Barcelona that scaled content is now a bigger problem than link spam.
Oct 4–6GoogleThe September 2026 spam update rolls through what looks like its final phase.
Oct 5OpenAIOpenAI says it will watermark ChatGPT's text for EU users, with the detector limited to approved researchers.
Oct 6GoogleSearch Console starts emailing site owners how many people picked them as a Preferred source.
Oct 6OpenAIChatGPT tests labels like “Best all-rounder” on product results.

The details, ranked by what they mean for your site:

Google put scaled AI content and fake authors in writing as its next target

Two Google documents changed on October 1, and they point the same way.

Google's guidance on AI-generated content now opens its accuracy section with a plain explanation of why it worries: generative models "don't retrieve facts, but predict a likely sequence of words", so their output can contain hallucinations. The conclusion follows from that: all AI-generated content should be manually fact-checked before publishing. Google spells out that the review covers the parts nobody looks at: <title> elements, meta descriptions, structured data and image alt text.

The same day, Google's guide to creating helpful content was rewritten around the main content of a page, which the Search Quality Evaluator Guidelines treat as one of the most important factors in page quality. Raters now judge it on four named attributes: effort, originality, talent or skill, and accuracy.

The effort section is the one to read twice. In Google's words, generating pages from feeds or using generative AI to produce large amounts of text "without manual oversight or curation represents little to no effort." And one line closes a popular loophole: "Attribution or giving credit to other sources doesn't replace the need for original effort." A page that summarises five other pages and links them all is still a page that added nothing.

The guide also added a line on authors. Fabricating creator profiles, with AI-generated headshots, made-up names or false credentials, is now described as "a form of deception".

What Google now counts as main content

The definition is wider than most site owners assume. Main content is any part of a page that directly helps it achieve its purpose, and Google lists five kinds:

primary text and media,
interactive features such as calculators, tools, games and site search,
user-generated contributions, when they serve the page's purpose,
tabbed or expanded sections,
page titles and headings.

Two of those matter for a typical WordPress site. The FAQ accordion and the tabbed product details you hid to keep the page tidy are main content, judged like everything else, so filler in them counts against you. And a working calculator or booking tool is quality in its own right.

There is reassurance in it too. Google says not all content needs specialised expertise, using someone sharing a better way to clear snow off a driveway as its example. The bar rises only where it should: on health, money and safety topics, content "must be highly accurate and consistent with established expert consensus."

Then Google said it out loud

At Search Central Live in Barcelona, which ran from September 30 to October 2, Gary Illyes told the room that scaled content is now becoming a bigger problem than link spam, and that Google is using AI to help catch AI-made spam. That is from attendee reporting by Search Engine Roundtable, not a Google transcript, and it says out loud what the documents say on paper.

Lily Ray put the pieces together on October 4 in a post titled Prediction: the next massive Google update is around the corner. Her list of likely targets is a list of tactics the AI visibility industry has been selling: self-promotional "best of" listicles that rank the author's own brand first, paid mentions on third-party sites, templated pages generated one per city, automated Reddit posting, and fake author profiles. She also counts the pace: one spam update in 2025, four so far in 2026.

Here is why a small business should care. None of this is aimed at people who wrote a thin page by hand. It is aimed at the shortcut: the agency that generated two hundred location pages for you last spring, the "expert" byline with a stock headshot, the meta descriptions a plugin wrote in bulk and nobody read.

Do this: look at the author box on your blog. If it names a person who does not exist, or shows a photo that is not them, fix it before Google decides what it means. Then spot-check your AI-written titles and meta descriptions for anything untrue, because Google now says those count too. If you bought pages in bulk, read five of them as a customer would and decide honestly which ones you would publish today.

Google's September 2026 spam update reached its final phase

The update announced on September 24 did not behave like the short spam updates earlier in the year. Barry Schwartz tracked waves on September 25 to 27 and on September 30, then what looks like a third and final phase between October 4 and 6. As of October 7 Google had not posted a completion notice.

The same week, Google gave the clearest answer it has offered on how long things take. Illyes's Barcelona slides, reported by attendees and not yet published by Google, put typical core update recovery at three to six months, and the slowest cases at six months to a year. A new URL is typically discovered in about 20 hours. A site move typically settles in one to three months.

That recovery figure matters more than any tactic. If your traffic dropped in late September, there is no fix that shows up next week, and anyone promising one is guessing.

Do this: mark September 24 and October 4 to 6 in your analytics. Once Google confirms the rollout is complete, compare the two weeks either side. If you lost ground, read the affected pages against Google's spam policies, starting with scaled content, then plan in months, not days.

Gemini started tagging the visitors it sends you

Gemini quietly began adding UTM parameters, the tags analytics tools use to name where a visit came from, to some of the links it sends people through. It surfaced on Reddit around October 1, Search Engine Journal reported it, and John Mueller acknowledged it. Google has not documented when the tags are added.

Our logs fill in the details Google did not. On the WordPress sites LovedByAI monitors, the first Gemini visit carrying a tag arrived late on October 4, and the tag reads exactly:

utm_source=gemini.google.com&utm_medium=ai-assistant

Two things are worth knowing. It is still a small minority: through October 6, about one Gemini visit in twenty-five carried the tags, and every one so far landed on a homepage. And every tagged visit also carried the normal gemini.google.com referrer, so if your analytics already counted Gemini by referrer, you were not missing these visitors. Observed in LovedByAI.

The reason this matters is attribution. Referrers get stripped by apps, privacy settings and redirects. A UTM tag survives all of that. ChatGPT has tagged its links for a long time, which is one reason it looks so dominant in analytics. The bigger blind spot is Google itself: a click from an AI Overview or from AI Mode arrives as plain google.com and lands in your organic line, so no report can split it out. Any chart showing ChatGPT with nine-tenths of "AI traffic" is really showing the assistants that identify themselves. If Gemini tags more of its links, its slice in your reports will grow without a single extra visitor.

Do this: in GA4, check that gemini.google.com with medium ai-assistant lands in the same channel as your other AI traffic, not in "Referral" or "Unassigned". If you build an AI channel group, match on both the source and the referrer, so a visit counts once whichever one survives.

ChatGPT's site: searches are reading your subdomains

ChatGPT has been leaning on the site: operator, which restricts a search to one domain, since the summer. On October 6, Lily Ray showed where that leads: the pages ChatGPT retrieved for a product-category question included a company's help centre on a subdomain. Help docs, forums, knowledge bases, anything that sits at something.yourdomain.com, are now part of how ChatGPT describes you.

Here is the part that bites small sites. Many WordPress businesses run a public staging copy at staging. or dev., with old prices, test products and placeholder text. If it is reachable, it is fair game.

Do this: list every subdomain you own. Put staging and development copies behind a password or a noindex, and read your help or documentation subdomain as if it were the answer an assistant is about to give, because it might be.

OpenAI will watermark ChatGPT's text in the EU

On October 5, OpenAI said it will add an invisible watermark to text that ChatGPT and Codex write for users in the European Union, rolling out over the following weeks. The EU AI Act's transparency rules, in force since August, require AI-generated content to be marked in a way machines can detect. Developers anywhere can switch it on through the API; it is off by default.

The watermark, which OpenAI calls textGrain, works by nudging the model's word choices into a pattern a detector can spot and a reader cannot. It travels with copy and paste. It does not survive much editing: in OpenAI's own tests, swapping a tenth of the words for synonyms cut detection from about 92% to 66%, and a quarter of the words cut it to 17%.

Two things keep this from being the scare it will be sold as. The detector is limited to approved researchers and organisations, so neither Google nor a competitor can run your pages through it. And OpenAI says itself that a watermark shows a passage passed through its model, not how much human judgement went into it. That is the same line Google drew in its guidance this week: the problem is unreviewed text at scale, not a tool.

Do this: nothing urgent outside the EU. If you publish AI-assisted text, the habit that protects you from both a watermark and Google's effort test is the same one: edit it properly, add what only you know, and put a real name on it.

More AI search news from October 2026

Cloudflare opened Pay Per Use on September 30, a beta that pays publishers when an AI product actually uses their content in an answer or recommendation, rather than when it crawls. Publishers opt in from the Cloudflare dashboard. No paying AI companies were named, so for now it is a mechanism without buyers.
Search Console now emails site owners how many people chose them as a Preferred source, the setting that shows a site more often in Top Stories and AI Overviews, reported October 6. If you publish news or regular articles, it is the first number Google has given you for it.
ChatGPT added virtual try-on for clothes and accessories on October 1, per OpenAI's release notes, along with Favorites and folders for saving products. On October 6 Brodie Clark spotted it testing labels like "Best all-rounder" on product results. ChatGPT is building a shop window, and the labels are its own judgement of your product, not yours.
Google's crawl rate documentation now explains the Retry-After header, updated October 6, so a 503 or 429 response can tell Googlebot when to come back. Nothing about crawling changed; it is useful if your host struggles under bot traffic.
The Search generative AI control page in Search Console now notifies the owner of a domain property when a child property overrides its setting, per Google's help page. Worth knowing if an agency or a previous developer verified part of your site separately.
Search Console's Performance report now lets you select several countries at once.
Google Search profiles, a page that pulls a creator's or publisher's accounts and posts into one place in Search and Discover, expanded to India and Canada on October 6, with Brazil next.
Google Maps' content policy now spells out that merchants cannot post ads in their listings or posts.
Nano Banana 2.1, Google's image model, is now inside AI Mode. Nothing for a site owner to do.

AI referrals and crawlers on small business websites in October 2026: what the logs say

These numbers come from the referral and server logs of the small business websites LovedByAI monitors. Six days is a short window, so every number here compares October 1 to 6 with the same six days of September, on the same sites, and should be read as provisional.

1 in 25Gemini visits arrived carrying the new UTM tags in their first three days, all on homepagesSmall business referral logs · LovedByAI · Oct 4–6 2026

−25%Gemini referral visits to the same sites, spread across many sites rather than oneSmall business referral logs · LovedByAI · same sites · Oct 1–6 vs Sep 1–6 2026, partial month

−42%how much less ChatGPT's live fetcher read the typical site. Nearly 9 in 10 sites with a real baseline were read lessSmall business server logs · ChatGPT-User fetches · sites with 20+ fetches in Sep 1–6 · Oct 1–6 2026, partial month

ChatGPT read far less and sent about the same

ChatGPT-User is the agent that fetches a page live because someone asked ChatGPT something. September's edition covered how it read each site less all month while sending more visitors. October picked up where September left off: the typical site was read 42% less than in the first six days of September, and the daily rate ran about 13% below the last ten days of September. It still reached nearly as many sites a day, about 4% fewer. Other assistants did not fall with it, which is how we know this is ChatGPT's behaviour and not a gap in our data.

Visits held up. On the same sites, ChatGPT's referral visits to the typical site were roughly flat, a small dip on the median site against a 42% drop in fetches. Fewer live reads and the same people sent is what you would expect from an assistant answering more from its stored copy of the web.

Gemini sent fewer visitors in the same week it started tagging them

Gemini's referral visits to the same sites fell by about a quarter. The drop is spread across many sites, so it is not one site losing a ranking. It is also six days, and Gemini's numbers have moved around before. The useful part is the coincidence of timing: if Gemini's tagging spreads, some of what you see in analytics over the next month will be relabelling, not growth, so compare like with like before you celebrate or worry.

Crawls are a means. Citations and visits are the end. If you want to know where your own site sits, the free GEO checker reads a live URL, our on-page GEO guide covers what to fix in order, and last month's edition covers what changed in September.