Last updated October 28, 2026
Between August and October 2026, our research team ran a controlled, two-wave study of how content watermarking affects search performance. In the first wave, from August 2 to August 24, we published and tracked 1,682 pieces of content across 139 websites in 4 B2B industries. Of those pieces, 63% were created with AI tools and therefore carried an invisible AI watermark, while the remaining 37% were produced without any AI involvement and thus could not be watermarked. In the second wave, from September 21 to October 16, we published 1,594 new pieces on the same 139 websites under the same method, 65% of them watermarked. All content in both waves was produced after August 2, 2026, the date by which the major AI models had formally introduced watermarking.
We measured fresh content in each wave, at the same age, rather than re-measuring the original pieces. This design separates a change in how search platforms treat watermarked content from the ordinary effects of content getting older and accumulating links.
Introduction of AI Watermarking By The Big 3 AI Platforms
As of August 2026, watermarking has become standard across the major AI platforms. Google, Anthropic, and OpenAI say they adopted these measures to meet new disclosure rules such as Article 50 of the EU Artificial Intelligence Act, and to limit the spread of undisclosed synthetic media.
Each AI platform watermarks in a slightly different way:
- Google’s SynthID embeds an invisible statistical signal into the word choices of Gemini’s text and into the pixels of its images
- OpenAI attaches C2PA content credentials, a cryptographically signed metadata standard, to ChatGPT’s image output, alongside a SynthID watermark
- Anthropic applies an invisible watermark to Claude’s text and C2PA metadata to the image files Claude generates
Each method embeds a marker a machine can detect (but a reader cannot). While text watermarks take the form of recognizable patterns in the characters and words generated, image watermarks alter pixels imperceptibly while attaching signed provenance metadata. Search engines and AI assistants can read these signals, which raises a practical question for publishers: Does watermarked content perform differently from un-watermarked content in search?
Detection has since become more accessible for media but not for text. On October 7, 2026, Google opened its SynthID Detector to the public, covering images, video, and audio from Google and partners including OpenAI. Text remains outside the public tool, and reading a text watermark still requires a key-based statistical test held by the model provider.
In the sections below, we report the ranking and citation results for watermarked and un-watermarked content across both study waves.
Search Performance of Watermarked vs Un-watermarked Content
In the following table, we summarize the impact of watermarking on the two metrics we studied, traditional Google rankings and citation rate on AI platforms, in each wave of the study.
Search Performance of Watermarked vs Un-watermarked Content, 2026
| Metric | Un-watermarked, August | Watermarked, August | Un-watermarked, October | Watermarked, October | Gap, August to October |
| Average Google Ranking (Position) | 6 | 11 | 6 | 8 | 5 positions to 2 |
| Average Citation Rate | 12% | 7% | 11% | 9% | 5 points to 2 |
- In our August wave, we found un-watermarked content ranked about five positions higher on Google. By October, the gap had narrowed to two positions, at 6 versus 8.
- The citation gap narrowed from five points to two, as watermarked content rose from 7% to 9% while un-watermarked content held near 12%.
- Un-watermarked performance stayed nearly flat across waves, so the narrowing came almost entirely from watermarked content improving. Watermarked content still underperformed in both waves.
A limitation of our study is that we did not control for content quality beyond our own professional standards; human-generated content may have received more thoughtfulness and judgment than AI-produced content, even when the AI content has undergone thorough human review. Because both waves were produced by the same teams under the same standards, this limitation applies equally to each and does not account for the change between them.
Average Google Ranking for Watermarked vs Un-watermarked Content
We define average Google ranking as a piece’s position for its target keyword, measured within three days of publication. In the chart below, we compare the two groups in each wave.
- In our August wave, watermarked content ranked roughly five positions lower on average, often the difference between the first and second page of results. By October, the gap usually left both groups on the first page.
- Because we measured within three days of publication in both waves, the gap reflects early ranking behavior, before longer-term signals such as backlinks accumulate.
- The gap appeared across all four B2B industries in August and narrowed in all four by October, which suggests neither the effect nor its decline was driven by a single vertical.
Average Citation Rate for Watermarked vs Un-watermarked Content
We define citation rate as the share of pieces cited for at least one of their target keywords when those keywords were entered into Google’s AI Overview, ChatGPT, and Claude. In the chart below, we compare the two groups in each wave.
- In our August wave, un-watermarked content was cited almost twice as often as watermarked content. By October, watermarked content was cited at more than four-fifths the rate of un-watermarked content.
- Citation rates were lower than ranking visibility for both groups in both waves, consistent with AI answer engines drawing from a narrower set of sources than the standard results page.
- The watermark is machine-readable, which raised the possibility in August that answer engines discount content they can identify as AI-generated. The October results suggest that discount has weakened, though our study measures the correlation rather than the mechanism.
In the table below, we break down citation rates by answer engine in each wave.
The Citation Rate by AI Answer Engine, 2026
| AI Answer Engine | Un-watermarked, August | Watermarked, August | Un-watermarked, October | Watermarked, October |
| Google AI Overview | 14% | 8% | 13% | 11% |
| ChatGPT | 11% | 7% | 11% | 9% |
| Claude | 10% | 6% | 10% | 7% |
Google’s AI Overview showed the largest change, with its gap narrowing from six points to two. Claude showed the smallest, narrowing from four points to three, which leaves it the answer engine with the widest remaining gap in our October data.
Why the Watermark Gap Narrowed
Our data suggests that search and AI platforms reduced how much weight they give watermark signals between August and October 2026. In the table below, we break down the ranking and citation gaps by industry in each wave.
The Watermark Gap by B2B Industry, 2026
| B2B Industry | Ranking Gap, August (Positions) | Ranking Gap, October (Positions) | Citation Gap, August (Points) | Citation Gap, October (Points) |
| B2B SaaS | 5 | 2 | 5 | 1 |
| Manufacturing | 4 | 2 | 4 | 2 |
| Financial Services | 6 | 3 | 6 | 3 |
| Healthcare | 5 | 2 | 5 | 2 |
Three features of the study point to a change on the platform side. First, each wave measured new content at the same age, within three days of publication, so the narrowing cannot be explained by pieces aging or accumulating backlinks. Second, the inputs held steady: the same 139 websites, the same content teams, and a near-identical mix of AI tools, with Claude producing roughly half of the watermarked content in both waves. Third, the gap narrowed in all four industries and on all three answer engines at once, a pattern more consistent with a shift in how platforms score content than with anything particular to a vertical or a tool.
No platform has announced a change in how it treats watermarked content, and our study measures correlation rather than mechanism. One plausible reading is that ranking and citation systems placed an early, blunt weight on machine-readable AI signals after the August 2 disclosure deadline, then shifted back toward the quality signals they have always relied on. Financial Services retained the widest gap in both waves, which may reflect stricter trust standards for content in regulated categories.
Requesting a Copy of This Report
If you’d like to request a PDF copy of this report or learn more about how our agency approaches content and search strategy, you can reach out here.
