SynthID AI Detector and AI Content Detection in 2026

If you publish AI-assisted pages, you have probably wondered whether Google can detect them and whether detection leads to a penalty. The published evidence answers both questions more narrowly than the usual SEO commentary does. This article covers what SynthID does, what Google’s policy says about AI-generated content for SEO, which risks you can measure on your own site, and where AI crawler access fits in.

What SynthID Does and Does Not Do

SynthID is a watermarking system from Google DeepMind, described on the SynthID page at Google DeepMind. It started with images and later extended to text, audio and video. The text version, SynthID Text, works at generation time. While the model picks each next token, it nudges the probabilities in a pattern that a matching detector can later test for statistically. The text reads normally, and the signal lives in the word choices. Google released SynthID Text as open source in 2024, so other developers can add it to their own models. Check the DeepMind page for the current scope before you cite any detail, because the product has changed several times.

A SynthID detector answers one question: did a model that embeds this watermark produce this text? A positive result means yes. A negative result means the detector found no SynthID signal. The text could come from a model that never embeds the watermark, from a Google model with watermarking switched off, or from a person. Most of the AI writing tools SEOs actually use do not embed SynthID, so a clean result tells you almost nothing about how a page was written.

The signal also has limits. Heavy rewriting, translation and paraphrasing weaken a text watermark. Short outputs and factual answers with little room for word variation carry less of it. A watermark says nothing about quality either, so a well-edited, accurate article and a careless one can carry the same signal.

General AI content detectors work differently. They have no embedded signal to look for. They guess from writing patterns such as predictability and sentence rhythm, and they return a probability. That guess breaks down on edited text, on non-native English and on plain technical prose, all of which can score as machine-written. A detector score is a vendor’s estimate of writing style, and Google’s ranking systems are a separate matter.

Google has not said that SynthID is a signal in Search ranking. No documentation, announcement or spam policy describes one, so a claim that Google reads the watermark and demotes you has no support. If that changes, it will appear in Search Central, and you can check there instead of trusting a screenshot of someone’s detector result.

You are really dealing with two questions. Whether a text can be identified as AI-generated is a technical question with a narrow answer. Whether a page holds up in Search depends on quality, accuracy and intent. Time spent on the first question does not answer the second.

Google’s Published Position on AI-Generated Content

Google’s position on AI-generated content for SEO fits in one sentence, and it appears in the Search Central page on using generative AI content: “Appropriate use of AI or automation is not against our guidelines. This means that it is not used to generate content primarily to manipulate search rankings, which is a violation of our spam policies.”

The same guidance says Google’s “focus on the quality of content, rather than how content is produced” is what has guided its ranking systems for years. Production method does not appear as a ranking criterion. Intent and usefulness do.

Enforcement sits in the spam policies. Google introduced the term scaled content abuse in March 2024 and defines it like this: “Scaled content abuse is when many pages are generated for the primary purpose of manipulating Search rankings and not helping users. This abusive practice is typically focused on creating large amounts of unoriginal content that provides little to no value to users, no matter how it’s created.”

The closing clause, “no matter how it’s created”, covers human writers too. A content farm paying freelancers to rewrite the same listicle 4,000 times falls under the same policy as a script calling a language model 4,000 times. One definition applies to both.

Three practical consequences follow from the published text.

• Google asks whether a page was made to help users or to capture rankings. A reviewed, edited, fact-checked draft that started with a model is within guidelines. A thousand templated pages with swapped city names are outside them, whoever typed the text.
• Quality gets judged on the output. The same Search Central guidance points to its helpful content and E-E-A-T material for what good looks like: original information, accuracy, evidence of first-hand experience and clear authorship.
• I found no official statement that Search reads watermarks or detector scores when ranking pages, and Search Console reports no AI content score. That is an absence of evidence and does not prove Search never will. Say so plainly if a client asks.

The assumption that detection leads to a penalty has no support in the policy text. Penalties and demotions in the published framework attach to spam behavior and low-value pages, and a detector flag is neither. A page that is thin, duplicated and written for the algorithm is exposed whether or not any tool can tell a model wrote it. A page that is accurate, specific and useful stays safe even when a tool says it looks like AI.

Google edits these pages, and wording shifts between revisions. Open both pages before you cite them to a client or paste them into an audit document, and note the access date.

Quality Risks You Can Measure

The policy asks whether a page helps people or exists to game rankings, so the risks worth managing are quality risks. Four of them matter, and none depends on a watermark.

Scaled content abuse comes first. The policy applies whether the pages come from a model, a freelancer pool or a template. A site that publishes 2,000 landing pages with swapped place names is exposed. A site that publishes 40 carefully edited articles drafted with an AI tool is not exposed by the policy’s wording. Volume combined with thin, repetitive pages is the trigger, and you can measure it on your own site.

Factual errors and invented sources come second. Language models produce confident numbers, quotes and citations that do not exist. A page that cites a study nobody can find, or states a wrong default value in a technical guide, loses reader trust first. Rankings follow slowly, through weaker engagement and fewer links. The only fix is a person verifying each claim before publication.

Missing experience is third. E-E-A-T asks whether the page shows first-hand knowledge. A generic model draft rarely contains a crawl you ran, a client migration that went wrong or a screenshot from your own Search Console. Pages that restate what the top ten results already say give a reader nothing new, and Google can see that gap without knowing how the text was written.

Opaque authorship and sourcing is fourth. Anonymous pages with no author, no publication or update date and no outbound links to primary sources give readers and quality systems little to evaluate. Compare a page that names its author, links the RFC or the Google documentation it relies on, and shows when it was last revised. The second page is easier to trust, whoever or whatever typed the first draft.

SynthID recognizes only text from models that embed its watermark, and heavy editing may erase the signal anyway. Build your process around the four risks above, because you can find, measure and fix those.

A good first move is to sort your pages by how they were produced. Pages made in bulk from a template, pages edited by a subject expert and pages with no named author each carry a different level of exposure. That sorting is the first step of the audit below.

A Practical Content Audit

The audit starts with a spreadsheet and a crawl. It needs no watermark detector, because every step looks at signals you can measure yourself.

1. Build a content inventory. Export every indexable URL from your CMS or sitemap and add four columns: publish date, author, how the page was produced (hand-written, AI-assisted, templated, bulk-generated), and organic clicks over the last 12 months. Most teams have never filled in the production column. Ask whoever ran each content project, because the CMS rarely records it. Sort by production method first, since bulk-generated batches are where scaled content abuse tends to start.

2. Crawl for templated writing. Run a full crawl with JavaScript rendering on and export titles, H1s, word counts and meta descriptions. Then look for patterns:

   • Hundreds of pages under a few hundred words, all on one URL pattern.
   • Titles that differ only by a city, product or keyword swapped into the same sentence.
   • H1s identical to the title tag across a whole directory.
   • One meta description repeated across a section.

   Open ten pages from each cluster and read them. If you can swap the topic noun and the page still reads the same, it gives a user nothing.

3. Check duplicate and near-duplicate signals. Exact duplicates show up in a crawler’s content hash column. Near-duplicates need a manual sample, because a template with 80 percent shared text and a different product name will not match. Compare those clusters against Search Console. Pages with impressions near zero and a “Crawled, currently not indexed” status are your first candidates for merging, rewriting or removal.

4. Verify authorship and sourcing. For every flagged page, check whether a named author exists, whether that author has a profile page, and whether claims link to primary sources. Then open the links. AI-assisted drafts often cite pages that do not exist or do not say what the sentence claims. Check any statistic, quote or date the same way, and remove the ones you cannot trace.

5. Check freshness markers. Compare the visible date, the date in structured data and the sitemap lastmod value. A page that shows “updated 2026” but has not changed in two years invites distrust.

6. Check image provenance. If you publish AI-generated images, look at whether your pipeline keeps C2PA or IPTC metadata or strips it on upload. The C2PA site hosts the specification. Support varies by platform, so test with a few files before assuming anything.

Sort the results into three piles: keep, rewrite with real expertise, and merge or remove. Pages in the second and third piles carry the risk, whatever tool made them.

AI Ready and Content Detection Answer Different Questions

The audit above covers content quality. A separate question is access, and teams often confuse it with detection. AI Ready looks at whether AI systems can reach your site. Content detection looks at whether anyone can tell that a machine wrote your text.

Crawlability is a set of access decisions you make. Your robots.txt rules decide which AI user agents may fetch your pages. An llms.txt file, the proposal described at llmstxt.org, gives language models a curated map of your content, and our generative engine optimization guide covers how that fits with AI search. Bot management in a CDN or firewall, such as the AI crawler controls in Cloudflare’s bot documentation, can allow or block traffic before robots.txt is ever read. OpenAI documents its own crawlers in its bots documentation, so you can write rules per bot instead of blocking everything.

Content detection sits at the other end. A watermark such as SynthID, or a classifier that scores text, looks at what a page says after it has been fetched. It has no view of your robots.txt, and your robots.txt has no effect on it. Blocking GPTBot does not make your articles look more human, and allowing every AI crawler does not make them look more machine-written.

The two meet in one place, which is how AI systems treat your site. A page that a crawler can never fetch cannot be cited in an AI answer, however good it is. A page that is fetched freely but is thin, templated and unsourced makes a weak citation candidate for reasons unrelated to any watermark.

Run the checks separately. Confirm that the bots you want can reach your key pages and that the ones you don’t want are blocked. Then audit quality on its own terms.

Seodisias is a free desktop crawler for Windows, macOS and Linux, and its AI Ready check covers the first job. It measures how accessible your site is to AI crawlers and agents. It does not detect whether your content was written by AI. To see which AI crawlers can reach your site today, run the AI Ready check on your domain.