Auditing Your Site for Risky GEO Tactics

The published evidence on citation rewriting points one way. C-SEO Bench (NeurIPS 2025) tested conversational SEO methods across several models and concluded that "most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking." The 252,000-trial factorial study (arXiv:2605.25517) found that formatting-only edits have little impact. A July 2026 survey of GEO research (arXiv:2607.14035) warns that citation-oriented rewrites can impair retrieval.

The original GEO paper is often quoted for gains, but it ran on a lab engine with sources already loaded into context. Even there, keyword stuffing lowered its own visibility metric (17.8 against a 19.5 baseline).

Two caveats keep this honest. These studies measured pages already inside the model's context, so they say nothing about whether anything helps a page get retrieved in the first place. They also predate the newest 2026 models. The safe reading: these tactics have no evidence of benefit and some evidence of harm. Google adds its own line in its AI optimization guide: "You don't need to write in a specific way just for generative AI search."

For the full argument, see whether rewriting content for AI can make visibility worse.

Appended quote blocks

A blockquote bolted onto the end of a page or section, often attributed to "an industry expert" or a vague title with no name. The quote restates the page and adds nothing. Red flag: the same quote structure appears on dozens of pages, or the person quoted cannot be traced.

Injected statistics

Numbers dropped into copy to look citable: "studies show 73% of buyers..." with no source, no date and no link. These are different from your own verifiable facts. A page that states your real price or your own measured figure gives an engine a sentence it can quote, which is the one positive on-page signal in the evidence. An invented or unsourced number is a factual liability. More on this in whether statistics and quotes help AI citations.

Keyword density spikes

One phrase repeated far more often than on comparable pages, or every variant of a query packed into headings. Google's guide says you don't have to worry about capturing every long-tail variation.

Authority boilerplate

Stock phrases such as "according to authoritative sources", "experts agree" or "as a leading authority in". They claim authority without naming any. They often arrive as a template block added by a plugin or a writing tool.

A separate, higher-severity class

Hidden text, instructions aimed at AI models, or content served only to AI user agents. This is adversarial content. Research on LLM search agents (arXiv:2606.16821) shows manipulated content can sway answers once retrieved, which is exactly why engines and Google's spam policies treat it as abuse. If you find it, remove it.

Say you run a B2B software site with 400 pages, half written or rewritten with AI tools in the last year. Here is a workable process.

1. Export a URL list. Pull every indexable URL from your sitemap. Tag each with its owner, publish date and whether it was rewritten for "AI optimization".
2. Fetch the raw HTML. Crawl each URL without running JavaScript and save the text. This is close to what most AI crawlers receive.
3. Search for quote blocks. Find every blockquote and every paragraph that opens with a quotation mark and closes with an attribution. Flag quotes with no named, traceable person.
4. Search for numbers. Extract every sentence containing a percentage or a large figure. Flag any without a named source, a link or a clear statement that it is your own data.
5. Search for boilerplate. Run a phrase list across the site: "according to authoritative sources", "experts agree", "studies show", "leading authority", "trusted by experts". Count hits per page.
6. Measure repetition. For each page, count how often its main target phrase appears, then compare against your site median for similar page types. No study sets a threshold, so pick one and apply it consistently.
7. Diff browser and crawler views. Request a sample of pages with a normal browser user agent and with AI crawler user agents. Any text that only one version shows is a cloaking flag. Also scan for text hidden with CSS or placed in attributes humans never see.
8. Log and grade. Record each flag with the URL, pattern and severity: hidden or cloaked content first, false or unsourced facts second, boilerplate and repetition third.

• No quote block without a named, traceable person and a reason to be there
• Every statistic has a named source, a date, or is clearly labelled as your own data
• No "studies show" or "experts agree" without the study or expert named
• No stock authority phrases repeated across templates
• Target phrase frequency in line with comparable pages on your site
• Headings written for readers, not stuffed with query variants
• No text hidden by CSS, zero-size fonts or off-screen positioning
• No instructions addressed to AI models anywhere in the HTML
• Same visible text served to browsers and to AI crawlers
• Key facts (price, plan names, product names) match across pages
• Page was not mass-generated alongside hundreds of near-identical pages

Scale matters too. Google's spam policies define scaled content abuse as many pages generated "for the primary purpose of manipulating search rankings and not helping users." In Otterly's 2026 experiment, two sites of roughly 1,000 AI-generated posts each were deindexed by Google with no manual action reported. If your audit finds hundreds of templated, thin pages, read whether optimizing for AI can hurt your Google rankings.

Fix in severity order.

• Hidden or cloaked content: delete it and serve one version to everyone.
• Unsourced statistics: source them, replace them with your own verifiable figure, or cut them. A wrong number is worse than no number.
• Quote blocks and boilerplate: delete them. If a real expert said something useful, integrate it with a name and context.
• Keyword spikes: rewrite for the reader. Say the thing once, clearly.

Do not expect a citation lift from the cleanup. The evidence supports removing risk, not a promised gain. The levers with stronger evidence are access levers: server-side rendering, per-bot robots.txt permission, firewall reachability and snippet controls. The comparison of on-page rewrites against access fixes lays this out. Once the content is clean, check that AI crawlers can actually reach it, for example with the AI Crawler Checker, and review snippet controls across the site so a template isn't suppressing pages by accident.

Run the full audit once, then repeat on a schedule tied to how content changes.

• Quarterly: full crawl with the phrase list, number extraction and repetition check.
• On every template or plugin change: boilerplate often ships sitewide from one component.
• Before and after any AI rewriting project: compare flag counts so a rewrite doesn't quietly add patterns.
• Monthly spot check: the browser versus crawler diff on your top pages.

Your own site is only half the exposure. Third-party pages that mention your brand can carry injected instructions too, so pair this with scanning third-party pages for injected instructions. For where this fits in a wider review, see what a real AI SEO audit checks.

Should I delete every statistic and quote on my site? No. The risk is unsourced or invented material added to look citable. A real, verifiable fact such as your price or your own measured figure is useful, and stating it plainly on a fetchable page gives an engine a sentence it can quote. Keep facts you can stand behind, attribute quotes to named people, and remove anything that exists only to bait a citation.

Is AI-written content itself a risk flag? Not on its own. Google's guidance says appropriate use of AI or automation is not against its guidelines. The risk is volume without value: many pages generated mainly to manipulate rankings, which Google's spam policies call scaled content abuse. Audit for the patterns and the scale, not for which tool wrote the draft.

Does the C-SEO Bench finding apply to the newest models? It is 2025 benchmark evidence and has not been replicated on every 2026 model. It also measured pages already inside the model's context. So it shows these rewrites lack evidence of benefit and can harm, but it does not prove content quality is irrelevant. As of October 2026, no published study shows these tactics working reliably across platforms.

Will removing these tactics improve my AI citations? Nobody can promise that. Citations depend on access, retrieval and model behaviour you don't control. Removing risky patterns lowers factual and spam risk. To improve your chances of being considered, check access first: rendering, robots.txt per bot, firewall reachability and index eligibility on each engine.

How is this different from checking for prompt injection? Prompt injection is hidden instructions aimed at AI models, which is adversarial and illegitimate. This audit covers it as the highest-severity flag on your own site. The other patterns, like boilerplate and keyword spikes, are usually well-intended but unsupported tactics. Injected instructions on third-party pages about your brand need a separate scan.