Can rewriting content for AI make my visibility worse?

Rewriting content for AI can make things worse, and the best-designed tests say so. The C-SEO Bench study (NeurIPS Datasets and Benchmarks, 2025) reported that "most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking."

Two other sources point the same way:

• A July 2026 critical survey of 45 GEO studies (arXiv:2607.14035) warns that "citation-oriented rewrites can impair retrieval."
• The original GEO paper (Aggarwal et al.) found keyword stuffing scored 17.8 on its Position-Adjusted Word Count metric, below the 19.5 baseline for unedited pages.

Google's own optimization guide adds that you don't need to write in a specific way for generative AI search. So the default move is to stop rewriting for machines and fix what blocks them from reading you.

Read the results with their limits. Almost every published GEO effect size was measured on documents already placed inside the model's context. No live search index was queried.

That matters because AI answers pass through four stages:

1. Parametric: what the model knows with no web access. Nobody changes this quickly.
2. Access: whether a crawler can fetch and store your page.
3. Retrieval: whether your page enters the model's context for a query.
4. Generation: which retrieved pages get used and cited.

Rewrite studies test stage 4 only. The survey's conclusion is narrow: already-retrieved content can change its citation, but no reviewed technique shows a stable, cross-platform causal effect on discoverability. C-SEO Bench also found gains shrink as more sites adopt the same tactics, which the authors called zero-sum. See what happens when every competitor does GEO.

The honest caveats: these are within-context nulls. They don't prove content quality is irrelevant, and they weren't replicated on the newest 2026 models. The often-quoted 40% GEO visibility figure comes from the same lab setup.

The tactics that tested null or negative share a pattern: they change how a page reads to a model, not what it says. Treat these as risk flags:

• Keyword stuffing or long-tail variant padding.
• Appended quotation blocks and injected statistics.
• "According to authoritative sources" boilerplate.
• Formatting-only edits. The 252,000-trial factorial (arXiv:2605.25517) found these have little impact.
• Whole-page LLM rewriting to sound "AI-friendly."

Volume adds a second risk. Google's spam policy treats generating many pages with AI tools without adding value as scaled content abuse. That hits your Google rankings, and losing the index removes you from AI Overviews too. The audit playbook for risky GEO tactics walks through finding these on an existing site.

Injected instructions or hidden text aimed at AI crawlers are a separate category. They are illegitimate, and you shouldn't use them under any framing.

Not every edit is a gamble. Two content factors hold up across sources.

Topical relevance. The survey and the 252,000-trial factorial both name topical relevance and list position as the biggest drivers of being cited. Relevance means covering the subject a buyer asks about, not matching every phrasing.

Specific, quotable facts. Stating a price, a number or a one-line definition on a fetchable page gives an engine a sentence it can quote. The factorial found explicit price information and a recent timestamp helped consistently. Say you run a hypothetical accounting firm: one plain sentence with your monthly fee beats three paragraphs of "value-driven solutions."

Write those facts for people first. If an edit makes the page clearer to a buyer, it's worth doing whether or not any AI cites it.

Access fixes have the strongest evidence in the field, and they work by removing blocks. Check these first, in order:

1. Rendering. Vercel found ChatGPT and Claude crawlers fetch JavaScript files but don't execute them. Compare your raw HTML with the rendered page and move key facts, like pricing, into the server-rendered HTML.
2. Robots.txt, per bot. OAI-SearchBot controls ChatGPT search. GPTBot controls training only. Check PerplexityBot and Claude-SearchBot too.
3. Firewall and CDN. A WAF can return 403 to a crawler your robots.txt allows. The AI Crawler Checker requests a page as each crawler and shows which ones get turned away.
4. Snippet controls. Look for stray noindex, nosnippet, max-snippet:0 or data-nosnippet left by a template.

The full evidence trade-off is in rewrites versus access fixes compared. None of this guarantees a citation. It makes sure nothing you control is stopping one.

Will rewriting for AI also hurt my Google rankings? It can if the rewrites are keyword stuffing or mass-produced pages without value, which Google's spam policies target. Google also says you don't need to write in a special way for its AI features, so there's no separate AI version to maintain. Editing for clarity and accuracy is safe. More detail is in [whether AI optimization can hurt Google rankings](/questions/ai-optimization-hurt-google-rankings).

What did C-SEO Bench find about traditional SEO? It found traditional SEO strategies significantly more effective than conversational rewrites. But its SEO arm simply moved the target document to position 1 in the model's context window. No index was searched and no ranking was earned. It shows that position dominates, not that classic SEO produces AI citations, and no publisher can invoke that lever directly.

Should I undo rewrites I've already made? Remove the ones on the risk list: stuffed keywords, injected statistics, appended quote blocks and authority boilerplate. Keep edits that made facts clearer or more specific for readers. The evidence doesn't support panic reverting of every change, only removing tactics that tested null or negative and that read badly to humans anyway.

Is rewriting ever measurably helpful? The only positive experimental signal for on-page content is stating specific facts, such as a price or a clear definition, on a page crawlers can fetch. Even that was measured on pages already in context. Nobody has published a test showing a rewrite helps a page get retrieved in the first place, so treat any promised uplift with suspicion.

How do I know if crawlers can read my content at all? Fetch a key page without JavaScript and see whether your main facts appear in the raw HTML. Then check robots.txt rules for each AI search bot and test whether your CDN or firewall blocks them. A [free GEO audit](/free/free-geo-audit) that checks access before accuracy is a practical starting point.