• AI visibility

Guide · AI search

Does keyword stuffing still work in AI search?

No. In every published test we could find, stuffing a page with repeated query keywords failed to raise its visibility in AI answers, and it often lowered it. What still counts is whether the page genuinely answers the question, and whether it ranks well enough to be read at all.

By Underneath Team Guide Updated

The short version

  1. In the original 2023 lab test (Aggarwal and colleagues), keyword stuffing cut a page’s share of the AI answer from 19.3 to 17.7; the best rewrites raised it 41%.
  2. On Perplexity, a live AI search engine, the same tactic performed 10% worse than leaving the page alone (same study).
  3. In a 2026 shopping test by Bagga and colleagues, AI models told to watch for manipulation flagged keyword-stuffed listings 84.5% to 100% of the time.
  4. In our study of ChatGPT’s local picks, a 10.6-point edge for keyword-stuffed business names shrank to 2.4 points with more data, too small to tell from chance.

Does keyword stuffing help a page appear in AI answers?

No. Controlled tests show it adds nothing or makes a page less visible in AI-written answers.

The first test came from Aggarwal and colleagues (opens in a new tab), who coined the term generative engine optimization (GEO) in 2023. They built a simulated AI search engine. Only the top 5 Google results for each question were fetched, and GPT-3.5 wrote a cited answer from them.

The researchers then rewrote one source at a time in nine ways and measured how much of the answer drew on it. Keyword stuffing meant adding more keywords from the question, as in classic search optimization. Its position-weighted share of the answer fell from a baseline of 19.3 to 17.7, as Martinez’s 2026 review (opens in a new tab) of the results sets out. The best rewrites, such as adding quotations or statistics, improved the same measure by 41%.

The team repeated a smaller test on Perplexity. There, keyword stuffing performed 10% worse than the baseline.

Has anyone repeated the test since?

Yes. Later tests on other AI models found the same weak and inconsistent result.

Wu and colleagues at Carnegie Mellon (opens in a new tab) re-ran the original rewrites in 2025 on an engine built on Google’s Gemini, across three sets of questions. Keyword stuffing lost ground on the original test questions and gained on open research questions. That is not a lift anyone could plan around.

Question set (Gemini engine)Untouched pagesKeyword-stuffed pages
Original GEO test questions19.4418.05
Open research questions20.1822.68
Shopping questions18.3219.17

Each number is a page’s position-weighted share of the answer, out of 100. The authors’ own learned rewriting method beat every simple tactic in the table.

A wider benchmark by Puerto and colleagues (opens in a new tab) tested ten popular rewriting tactics, though not keyword stuffing itself, across six kinds of content and four AI models. In their main experiment, out of 54 combinations of tactic and content type, only three produced a reliable gain, and many rewrites pushed pages down. Moving a page higher in the list the AI received beat any rewrite.

Why doesn’t repetition work when AI writes the answer?

AI assistants read for meaning and write their own searches, so repeating a phrase gives them nothing new.

Classic search engines grew up matching the words on a page to the words in a query. The 2023 authors argued that AI engines are not limited to that kind of matching, because the model reads the whole page and the whole question.

Our hidden searches study shows how far the assistants move from the buyer’s wording. On 80 buyer questions, ChatGPT ran a mean of 3.7 searches per answer before replying. None of the 509 searches repeated the user’s question word for word. A page tuned to one exact phrase is tuned for a search the assistant may never run.

Once pages are retrieved, the AI picks what it can use. Martinez’s review of 45 studies rates the support for keyword stuffing as “null or negative” across multiple benchmarks, and its advice is simply to avoid it. The best-supported levers in that review are relevance to the question and a page’s position among the results the AI reads. We gather the rest in our guide to research-backed GEO practices.

Do keywords still matter at all?

Yes, in the plain sense: a page must use the words and cover the topic the buyer is asking about.

In a 2026 controlled test by Vishwakarma and colleagues (opens in a new tab) at Sprinklr, a software vendor, six AI models repeatedly chose between two versions of the same page. There were 252,000 trials in all. Being on topic was one of four factors that decided citation in all six models. A page missing key terms that a competing page used also lost. That factor was among 11 of the 18 factors (61%) that mattered in at least four models. Formatting changes alone made no difference.

The shopping test by Bagga and colleagues (opens in a new tab) points the same way. When software searched for the best instructions for rewriting product listings, the winners converged on one playbook. It included working in relevant keywords and synonyms, and it explicitly warned against repeated keywords and unnatural phrasing.

So the line runs between coverage and repetition. Naming the product, the problem and the terms buyers use helps an AI match a page to a question. Saying the same phrase again and again does not. Whether clearer, plainer prose helps instead is covered in our guide on readable writing and AI answers.

Does stuffing a business name with keywords help in ChatGPT?

Not reliably. Our own data showed an early edge that disappeared once we collected more answers.

Many local businesses add service and city words to their Google Maps name, such as “Business Name - Emergency Plumber - City”. ChatGPT copies those names faithfully: in our local recommendations study, all 115 keyword-stuffed names were reproduced verbatim on 26 September 2026.

Copying a name is not the same as favoring it. Our local picks study compared 898 ChatGPT answers with the Google Maps top 20 for 120 local searches in four countries. On the first day, stuffed names showed a 10.6-point advantage. It held up in only two of eight sets of answers, and after adjusting for rank and reviews it shrank to 2.4 points, within the range of chance.

Having more reviews than local rivals, by contrast, helped in all eight sets. The study also notes that Google’s guidelines do not allow keyword-stuffed business names, and a profile that breaks them can be suspended.

Is there a risk in stuffing, beyond wasted effort?

Yes. AI systems can read keyword-stuffed text as a sign of manipulation and rank it lower.

In the shopping test, each AI model ranking products received one short instruction to watch for manipulative descriptions. The researchers used a fixed set of 200 shopping questions. Listings rewritten to repeat category keywords were flagged as questionable between 84.5% and 100% of the time across five AI models. They also fell in the rankings on every model.

That warning instruction was the researchers’ own, not a known feature of ChatGPT or Google. Still, it shows how cheaply an AI can spot the tactic. The strongest models in the test, GPT-5 and Claude, gave manipulative rewrites little lift whether or not they flagged them.

What should you do about it?

Stop paying for keyword density and put the budget into pages that answer buyer questions completely.

  1. Remove keyword-density targets from content briefs and vendor scopes. No published test shows they help in AI answers.
  2. Make sure each important page names the product, the problem and the words buyers actually use, naturally and in context.
  3. Add what AI answers can reuse: prices, specifications, comparisons, dates and evidence for claims. These won in the controlled tests above.
  4. Keep investing in search ranking. In the benchmark tests, a page’s position among the results the AI read mattered more than any rewrite.
  5. Keep your Google Business Profile name as your real business name, and put the effort into earning reviews instead.
  6. Measure AI visibility directly, across several assistants and repeated runs, rather than assuming search tactics carry over.

If you want help setting up that measurement, see our generative engine optimization service.

What does the research not tell us yet?

The evidence is consistent but narrow, and most of it comes from lab setups rather than live search products.

  • Most tests hand the AI a fixed set of pages. None shows what keyword stuffing does to a page’s chance of being found in the first place.
  • Only the 2023 study checked a live engine, Perplexity, and AI products have changed many times since.
  • None of these tests measured what stuffing does to Google rankings, which still feed many AI answers. The 2023 authors say so themselves.
  • The manipulation-flagging result depends on a warning the researchers added. Live assistants may be stricter or more lenient.
  • Our local finding covers ChatGPT, four countries and two days in September 2026.

Frequently asked questions

Is keyword stuffing bad for ChatGPT visibility?

There is no evidence that it helps, and some that it hurts. In controlled tests, stuffed pages lost share of the answer, and AI models told to look for manipulation flagged stuffed listings most of the time.

Should I still do keyword research for AI search?

Yes, to learn the words and questions buyers use, not to repeat them. In the 252,000-trial test, pages missing a buyer’s key terms lost to pages that had them, while formatting tricks made no difference.

Does keyword stuffing work in Google AI Overviews?

No published study has tested it on AI Overviews directly. The closest evidence, from simulated engines fed with Google results and from an engine built on Gemini, shows no dependable gain.

What works better than keyword stuffing for AI answers?

Relevance, ranking and reusable evidence. In the 2023 lab test, adding quotations or statistics raised a page’s share of the answer by up to 41%. That gain applied only to pages the engine had already been given.

Sources

Free strategy call

Some questions are easier to answer about your own business.

Bring the one that matters most. On a free 30-minute call we’ll take a first look at it and send you a short written read afterward.

Ask about my businessHow we work