LLM Optimization: Which Tactics Have Proof Behind Them

LLM optimization is the work of getting engines like ChatGPT, Perplexity and Gemini to find your pages, cite them and name your business when a buyer asks a question you can answer.

The tactics sold under that name rest on very different amounts of proof. In Pew Research Center's survey of 5,119 U.S. adults, fielded February 17 to 23, 2026, 44% said they use ChatGPT, and about four in ten said they use chatbots to search for information. That national figure can't tell you how your own buyers research. It does mean a wrong guess has a price. You can pay for tactics with nothing behind them, and another firm can still be the name ChatGPT gives when someone asks who to hire.

Key points

• Of the tactics in the table below, only one is backed by OpenAI's own documentation: letting OAI-SearchBot, its search crawler, read your site.
• The lab study behind "add quotes and statistics" tested a GPT-3.5 Turbo engine and Perplexity, not the current ChatGPT, and one of its top methods lifted fifth-ranked pages while lowering the top-ranked page.
• You can run an exploratory test of one page in ChatGPT yourself before you pay anyone, as long as you log every run the same way.

What is LLM optimization?

LLM optimization (LLMO) is the practice of making your business easy for large language model engines, such as ChatGPT, Perplexity, Gemini and Claude, to find, cite and name in their answers. It covers whether their crawlers can read your site, what your pages say, and what other sites say about you. You measure it in the answers themselves.

Engineers use the same term for making models run faster or cheaper. This article covers the marketing meaning.

Which LLM optimization tactics have evidence behind them?

Start with the first row, because there the engine's owner states the effect. OpenAI's crawler documentation says OAI-SearchBot "is used to surface websites in search results in ChatGPT's search features." Blocking GPTBot, the crawler OpenAI uses for model training, is a separate choice ("Each setting is independent of the others"), so you can opt out of training and still let ChatGPT search read your pages. Our five-check teardown covers the robots.txt and firewall checks.

What did the "add quotes and statistics" study actually test?

Much of the advice to add statistics, quotes and citations traces back to the GEO paper by Aggarwal and colleagues, last revised in June 2024. Its abstract says GEO "can boost visibility by up to 40% in generative engine responses." Four details from the paper change what that number means for your site.

Which engines did it test?

The main experiments used the authors' own test engine, where "The answer is then generated by the gpt3.5-turbo model." They also ran the methods on Perplexity, but uploaded the source text as files and had Perplexity answer only from those files. That test didn't measure whether Perplexity finds a live business website. In that setup, Quotation Addition did best, with a 22% improvement over the baseline, and keyword stuffing did 10% worse than the unedited text.

What did it measure?

The 30% to 40% gain for Cite Sources, Quotation Addition and Statistics Addition was on one metric, Position-Adjusted Word Count, which the paper describes as considering word count and citation position in the engine's response. It tells you how much of an answer drew on a source and how early. It doesn't tell you whether a buyer saw your name or clicked.

Were the added statistics real?

Not necessarily. In the code the authors published, the prompt for the statistics method asks a language model to add statistics "even if hypothetical." Don't copy that. A made-up figure on your page is a false claim to your buyers.

Did a page's starting rank change the result?

Yes. In the paper's own engine, the Cite Sources method raised visibility 115.1% for websites ranked fifth in search results, and lowered it 30.3% on average for the top-ranked website. In that test, the method that lifted a fifth-place page cost the page already in first.

Real figures from named sources, with dates, still help the person reading your page. None of the sources here show that adding them gets you cited in ChatGPT today, so test one page before you rewrite them all.

Do mentions on other sites help you show up?

They correlate highly with visibility, though Ahrefs' data can't show that they cause it. Ahrefs studied 75,000 brands across ChatGPT, Google AI Mode and Google AI Overviews, in a study published in December 2025. YouTube mentions had the strongest correlation with AI visibility, about 0.737, and branded web mentions correlated at 0.66 to 0.71. ChatGPT showed the weakest correlations of the three engines for almost every traditional brand authority signal, with branded search volume at 0.352 and Domain Rating at 0.266.

Two limits matter for a local firm. Ahrefs kept only domains with a Domain Rating above 40 whose top keyword drew at least 800 monthly searches, so a smaller local site may sit outside the sample. And Ahrefs repeats the usual disclaimer that correlation isn't causation.

Even so, read what your industry directories, local news coverage and YouTube results say about you, and correct anything wrong. A wrong description your buyers read is worth fixing for its own sake.

How do you test an LLM optimization change in ChatGPT?

You'll end with an exploratory read on one question and one page, before you change your whole site. It can point you to a change worth keeping. It can't prove the change caused a shift, because ChatGPT's answers vary from run to run and you don't control what it retrieves.

1. Pick one buyer question and the page that should answer it. Hypothetical example: a small accounting firm picks "What does it cost to have an accountant file a business tax extension?" and its tax extension service page.
2. Confirm your robots.txt allows OAI-SearchBot.
3. Ask the question in a fresh chat with search on, with the same account, model and location each time. For every run, log the date, the exact prompt, the full answer and every cited URL.
4. Score each run twice. Named means your business name appears in the answer text. Cited means a link to your own domain appears among the sources. A run can name you and still cite a directory instead.
5. Make one change to the page, such as replacing a vague line with a dated figure from a named, linked source.
6. Look for signs ChatGPT has seen the new version. If your host shows server logs, check for an OAI-SearchBot visit to that page after the edit. If you can't confirm one, mark the comparison as uncertain.
7. Rerun the same question the same number of times and compare.

Hypothetical example: before the edit, the firm is named in 2 of 10 runs and cited in 1. After it, the firm is cited in 3 of 10. Treat that as a reason to retest, not a result. Ten runs is a small sample, and anything else that changed online in those weeks could explain the difference. For a fuller logging sheet, see how to test what ChatGPT cites.

Is LLM optimization worth paying for?

Run the crawler check and the one-page test first. They cost you time, not a fee, and they show whether you have a problem worth paying to fix. Then hold any proposal against the table above and ask which row each tactic sits in. Be wary of anyone who promises a citation, since no vendor controls which page ChatGPT cites, or who plans to add figures they can't source. If your buyers rarely compare options before they call, the robots.txt check alone may be enough for now.

Before you sign off on any LLM optimization work, read Vector Index's guide to LLM SEO and LLM optimization. It shows which kinds of questions get answers with cited sources and which don't, so you can decide which engines and questions deserve your time. From there you can ask Vector Index for a snapshot of where your business stands now.

FAQ

Is LLM optimization the same as SEO?

They overlap. SEO aims to rank your pages in search results. LLM optimization aims to get your business named or cited inside an answer, and for ChatGPT search it starts with letting OAI-SearchBot read your pages.

Does LLM optimization work the same in Perplexity, Gemini and Claude?

Not necessarily. Neither study above reported results for Gemini or Claude, so run your test in each engine your buyers use.

Does an llms.txt file help ChatGPT cite my site?

None of the sources in this article tested it, so there's no evidence here either way. Our sister site explains what llms.txt is if you're weighing one.

How long does an LLM optimization change take to show up?

None of the studies here measured how long a page edit takes to appear in ChatGPT answers. OpenAI does publish one timing. After a robots.txt update, it can take about 24 hours for its search systems to adjust.