AI Content and AI Slop: What Google Actually Says and How to Audit Your Site

Picture a software company that signs up for an AI content platform in January. By July its blog has grown from 80 posts to 2,300, with a comparison page for every pair of competitors and a glossary entry per term. Traffic doubles, then falls, and within a year it's below where it started. Someone pastes ten posts into an AI detector, gets "94% AI", and the meeting becomes an argument about whether Google penalises AI content. The company is a composite; the pattern isn't.

That's the wrong argument. Google's documents say the tool isn't the problem, and in 2026 they spell out what Google expects instead. Below: what Google says, what practitioners have seen, how to use AI well and how to audit for slop, including what DeepSpider.ai can and can't show you.

In short
Google doesn't demote pages for being AI-generated. It acts against content that adds little value, especially at scale, however it was made. Its 2026 documentation sets the bar: effort, originality, accuracy, real authors, and a manual fact-check of everything AI produced, metadata included. Use AI for research, structure and editing, and audit by page type rather than by detector score.

What "AI slop" means

Merriam-Webster chose slop as its 2025 Word of the Year and defines it as "digital content of low quality that is produced usually in quantity by means of artificial intelligence." Both parts matter for search: low quality, and quantity.

The Google documentation cited here doesn't use the word, but its spam policy on scaled content abuse describes the same thing: "many pages are generated for the primary purpose of manipulating search rankings and not helping users", typically "large amounts of unoriginal content that provides little to no value to users, no matter how it's created." The first example is "Using generative AI tools or other similar tools to generate many pages without adding value for users".

"No matter how it's created" cuts both ways. Freelancers rewording competitors' articles can produce scaled content abuse without AI, and AI-assisted content isn't suspect by default. Google said so in 2023: "Appropriate use of AI or automation is not against our guidelines", and "Using AI doesn't give content any special gains." Its quality rater guidelines add that "the use of Generative AI tools alone does not determine the level of effort or Page Quality rating."

What Google's documents now say

Google's position hasn't reversed in 2026; it has become more specific.

One caveat from the generative AI page: raters' "ratings don't directly influence ranking". The guidelines describe what Google's systems aim for; they aren't a scoring formula.

Fact-check everything, including the metadata

On 1 October 2026 Google updated its generative AI guidance to say that "generative models don't retrieve facts, but predict a likely sequence of words", and that "It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing." The review covers titles, meta descriptions, structured data and image alt text, because they can appear in search results.

Effort, originality and accuracy are spelled out

Matt G. Southern of Search Engine Journal compared the version of Google's people-first content page dated 1 October with a copy saved on 8 June and noted two additions (the page has changed again since). The first is a section on main content listing four attributes raters are trained to evaluate: effort, originality, talent or skill, and accuracy. Its definition of effort: "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." Originality means information or perspectives "that aren't already available on other websites." On topics that could significantly affect people's lives, content "must be highly accurate and consistent with established expert consensus."

Invented authors are deception

The second addition says: "Fabricating creator profiles (such as by using AI-generated headshots, made-up names, or false credentials to make content appear as if it was written by human experts) is a form of deception."

On disclosure, the page suggests explaining how content was made "when it would be reasonably expected."

Commodity and non-commodity content

In May 2026 Google published a guide to optimising for its generative AI features and defined what it calls commodity content. "Commodity content (for example, something like '7 Tips for First-Time Homebuyers') is often based on common knowledge, which could originate from anyone". Non-commodity content, such as "Why We Waived the Inspection & Saved Money: A Look Inside the Sewer Line", offers "unique expert or experienced takes".

It adds that pages for every variation of a search, including "fan-out queries", made primarily to manipulate rankings or AI answers violate the scaled content abuse policy, and that "a high quantity of pages doesn't make a website higher quality".

Nacho Mascort turns this into a test: "If you can swap your brand for a competitor's and the page is still worth the same, it's commodity."

What practitioners have seen

Google rarely says what an update targeted. People who analyse affected sites fill part of that gap, but their findings are observations, not proof of how ranking works.

Fast growth, then a steep fall

Lily Ray tracked more than 220 sites and subfolders publicly named as customers of AI content platforms. Using Ahrefs and Sistrix estimates, she found that "54% lost 30% or more of their peak organic traffic", 39% lost half or more and 22% lost 75% or more. Among the declines the shape repeated: pages grew for six to twelve months, traffic peaked within roughly three to six months of the content peak, then fell steeply within the following year, often below where it started.

She is careful: "these are third-party estimates, not first-party analytics", and "I am not asserting that any AI content tool directly caused any traffic outcome". Her sample comes from customer stories that AI content platforms published, so it isn't random. More useful than the percentages is her list of eight recurring patterns, including comparison pages at scale, "What is X" glossaries, "Best X for Y" lists, programmatic location and language pages, and FAQ farms with one question per URL.

More spam updates, and losses in AI answers

John Campbell's recap of Google's Search Central Live event in Barcelona (October 2026) reports Google Search Quality Analyst Duy Nguyen explaining that "there have been more spam updates recently because there is so much more new content." Google's Search Status Dashboard lists four spam updates in 2026 (March, June, August and September, the last still rolling out as we write), against one in 2025. Glenn Gabe documented blinded cases from the August 2026 spam update. A site whose programmatic pages also carried chunks of AI-generated text, in an "ultra-YMYL niche", lost rankings for over 200,000 queries. On another, "the spam update yielded impact across sections (site-level)." A site that's hit "can drop across surfaces, including AI Overviews and AI Mode", Gabe adds.

A human pass isn't enough

Asked on Reddit whether rewriting unindexed AI pages by hand would help, John Mueller replied, as reported by Search Engine Roundtable: "I wouldn't think about it as AI or not, but about the value that the site adds to the web. Just rewriting AI content by a human won't change that, it won't make it authentic."

Editing sentences doesn't add what a page lacks. Mascort describes how sites drift into the problem: "A page type that works on 100 pages gets launched on 10,000. The last 9,000 have less data, less demand and more filler text."

Weak sections can drag down the rest

At Softonic, Mascort's team moved translated versions from subdomains onto the main domain. A core update then cut traffic by 20-something percent, and they concluded the weaker translations were lowering the quality of the whole site. Once the languages had their own domains, the main domain recovered at the next core update.

What nobody outside Google knows
Which signals Google uses to recognise scaled content isn't public. Google fights spam with automated systems such as SpamBrain, and the Barcelona recap reports that its team "is now using AI to help catch more of it". That's spotting spam patterns, which is different from labelling who wrote a page; in the recap's words, "the point is the quality, not who wrote it." No Google document we reviewed mentions an AI-authorship detector for ranking.

Why a detector score is the wrong audit

AI detectors answer a question Google says doesn't decide quality. In a 2025 study, Ahrefs, which sells a detector, ran 600,000 top-20 pages through it. Only 4.6% came out as "pure AI" and 81.9% as a mix, which Ahrefs expects because people also use AI for spelling, grammar and titles. The correlation between AI share and position "was 0.011, effectively zero", though "Purely AI content rarely reaches the #1 position". Ahrefs also says detectors "deal in probabilities, not certainty", and Gabe pairs detector scores with his own review of the pages.

The rater guidelines go further: raters should give the lowest rating "when you strongly suspect scaled content abuse after looking at several pages on the website", even if they're "unsure of the method of creation". A human-written commodity page fails that test, and an AI-assisted page with your own data can pass it. A detector percentage is like an issue count without the evidence behind it, the problem we described in why technical SEO audits need context, not just issue counts.

When to leave it alone
An AI-assisted page that's accurate, specific and reviewed doesn't need rewriting to "sound human". Repeated text that does a job, such as shipping terms, legal notices or product specifications, isn't slop because it repeats. Location pages for places you really serve, with real local details, aren't slop just for being templated.

How to use AI without producing slop

Good workflows decide what you can add first, give AI the work that doesn't need your knowledge, and keep a person responsible for every claim.

Decide what only you can add, before you prompt

Despina Gavoyannis of Ahrefs calls repeating what's already out there one of the biggest mistakes in AI content. Her fixes: your own experience, small experiments with real data, expert interviews, and analysis that stays yours. Marie Haynes is blunter: "if your SEO company can use AI to create content on your topics, then it's likely not original, insightful and substantially more helpful than what currently exists". If you can't name what a page will add, don't publish it.

Use AI where it saves time without replacing knowledge

Edwin Toonen of Yoast: "It's fine to use AI, but use it to help you structure content or brainstorm, and don't publish anything directly." Google's guidance points the same way: AI "can be particularly useful when researching a topic, and to add structure to original content."

Fact-check claims, metadata and markup

Keep a source note beside every number, name, date and quote, and check each one. Compare AI-written alt text with the image, and check that structured data values such as prices, dates, ratings and author match the visible page. On health, finance and safety topics, have a qualified person sign off the claims.

Put a real person behind it, and say how it was made

Use real authors with true credentials, or an honest team byline, and say how a page was made where readers would wonder. Toonen suggests: "This article was drafted using generative AI and reviewed by [Editor Name], [Job Title] at [Company Name]." A label doesn't make a weak page good, though.

Publish less

Haynes warns that covering every fan-out query at scale means "you run the risk of facing a scaled content penalty." She suspects, but says she can't yet prove, that the June 2026 spam update hit sites creating commodity content at scale. Fewer pages, each with a reason to exist, are easier to defend.

How to audit a site for AI slop

An audit for slop is mostly an audit of page types. A single weak article rarely matters. A template copied across thousands of URLs does.

Step 1: group URLs by page type

Start with Mascort's advice: "Group URLs by page type and look at traffic, indexed percentage and growth for each group." Sort URLs from sitemaps or a crawl by folder or template, date their growth with CMS publish dates, honest sitemap lastmod values or two crawls, and compare it with what your team could realistically write and check. Volume alone isn't the problem: Patrick Stox calls scale "a prerequisite, not the offense".

Step 2: check what Google already decided

First check Security & Manual Actions > Manual actions in Search Console. Besides automated systems, Google uses human review that can end in a manual action. As Stox explains, a scaled content notice needs a reconsideration request after cleanup; a drop without one is algorithmic, so you fix the site and wait.

Then use the Page indexing report. It doesn't group by template, so submit a separate XML sitemap per page type and filter by sitemap to see each template's indexed share and "Crawled - currently not indexed" count. Mascort calls page types with a lot of that status "your first alarm"; Haynes: "In almost every case I've examined these pages have quality issues." The status only means Google hasn't indexed a crawled page, at least for now, so it isn't proof. But a template where most URLs sit there is telling you something, and on large sites those URLs take crawling away from pages you care about, as explained in your crawl budget is now shared with every Google crawler.

Step 3: look for the traces slop leaves

• Leftover assistant phrasing such as "As an AI language model", which the rater guidelines mention.
• Text that stops mid-sentence, or headings with nothing useful under them.
• The same paragraph on every location or product page, with only the name changed.
• Duplicate titles, descriptions and headings across a template, placeholder text, titles like "Untitled".
• Author pages for people who don't exist, or AI-generated headshots.
• Dates changed without real changes, which Google lists as a warning sign.

Step 4: read a sample like a rater

Open ten pages from each page type and read them. Ask Lily Ray's questions: could a competitor publish this tomorrow with the same prompt? Would you be comfortable if Google, a journalist or your customers saw every URL in this folder? Is there "any first-party data, expertise, or original perspective on this page that isn't available on the first ten results"?

Step 5: improve, consolidate or remove

Mascort: "What you can't turn into non-commodity, merge with other pages or delete." The spam policy adds: "If you're hosting such content on your site, exclude it from Search." Lily Ray's tool data suggests many brands are "removing, redirecting, or 410'ing" pages that had featured in their vendors' case studies.

Before removing anything, check each page's clicks, conversions and external links: pages with any of those are usually worth redirecting. Stox recommends a 301 "when there's a better, relevant destination" and 404 or 410 only when a page has no value and no equivalent.

Then measure in months: Mascort expects changes to show "with luck, in 3 to 6 months".

For page-by-page decisions, see our guide to content decay and when to update, rewrite, merge or leave a page. After a merge, update internal links that point to removed URLs and link the surviving page from related pages, as described in when a page has too few internal links.

How DeepSpider.ai helps

DeepSpider doesn't tell you whether AI wrote your pages, and that's deliberate. It shows patterns a person should review, with the evidence beside them.

The AI Content Detection panel (labelled Beta) sits on the audit Overview in Guided and Professional views and runs after crawling. Its three signals are an Assistant-style statement (fixed phrases such as "as an AI language model", English pages only), a Repeated paragraph in saved text on one page, and a Paragraph also found on another audited page: an identical paragraph of 150 or more characters on two pages, with a link to the other one. A paragraph where only a place name changes doesn't match. The panel always shows "AI authorship: Not estimated" and doesn't affect Website Health Score.

The duplicate checks and Vague page titles are Warnings that count towards Website Health Score. Near-duplicate main content, Placeholder text in main content and the author check are Needs review and don't. Each check comes with How to fix it guidance.

For scale, the Content growth since previous audit summary on the audit Overview (Professional view) and the Content changes tab in Compare Audits show new, expanded and shortened pages and the words added between two saved audits. They count; they don't judge, and a page new to the audit isn't necessarily newly published.

With Search Console data, Indexable pages without Google impressions lists indexable audited pages with no impressions in the period; no impressions doesn't prove a page isn't indexed. It needs the complete page table, so an uploaded interface export, capped at 1,000 rows, leaves larger sites Not evaluated unless Search Console is connected. A connection also adds Audited pages Google has not indexed, grouped by Google's reason, including "Crawled, currently not indexed".

DeepSpider gives no AI probability, doesn't compare your text with the wider web, doesn't check facts and can't decide whether a site breaks Google's spam policy. It reads source HTML without running JavaScript, and AI Content Detection reviews up to 2,000 HTML pages per audit and the first 4,000 characters of each page's saved text. No finding means no matching pattern, not that the content is good.

Frequently asked questions

Does Google penalise AI-generated content?

Not for being AI-generated. Google says appropriate use of AI "is not against our guidelines" and acts against low-value content produced at scale, whoever or whatever made it. Unreviewed bulk AI text, though, is Google's own example of little to no effort.

Do I have to label AI content?

Not for web search. Google suggests a disclosure where readers would reasonably ask how a page was made. Shopping is the exception: Google Merchant Center requires AI-generated product titles and descriptions to be labelled, and AI images to carry IPTC metadata. Invented authors are a different matter: Google calls them deception.

Is programmatic SEO always scaled content abuse?

No. Pages built from real data you hold, such as product details or local information, can be useful. If each page only swaps a keyword, it's at risk.

How long does recovery take?

Months. Gabe says "typically at least several months", and a site that keeps the same content may not recover at all.

The short version

Slop is low-value content at scale, and Google treats it the same whether people or models wrote it. Its 2026 documentation asks for effort, original information, accuracy, real authors and a manual fact-check of everything AI touches, metadata included. Audit by page type, check for manual actions and unindexed templates, read samples like a rater, then improve, merge or remove.

Sources