What works and what doesn't AI visibility tactics sorted by evidence: third-party proof, reviews and live search work; llms.txt alone, schema alone and directory blasts do not. By Dasha. Updated November 1 2026. https://heardofyou.io/guides/what-works-and-what-doesnt Short version. What works: third-party mentions, honest reviews where each assistant looks, a place in its live search, a clear answer up top, real update dates. What does not work, or cannot be sold honestly: llms.txt on its own, schema on its own, keyword stuffing, bulk directory submissions and hidden text. The evidence for each is below, with vendor studies labelled. What works 1. Get named on pages you do not control AI answers lean on earned and third-party media. For US software questions, 72.7% of cited sources were earned media. A normal Google results page: 45.4% (Chen et al., University of Toronto). Across 207 million citations, over 84% were earned or third-party and 0.3% paid (Muck Rack, vendor). "Best X" lists alone were 43.8% of the page types ChatGPT cited, across 26,283 source URLs. Recently updated lists were most prominent. Appearing high on several lists correlated with being recommended (Ahrefs, vendor, correlational). Which pages? Reddit and YouTube are in every engine's top five cited domains. G2 is in Perplexity's (Peec, vendor). Across 75,000 brands, YouTube mentions had the strongest correlation with AI visibility (about 0.737). Branded web mentions came next (Ahrefs, vendor). On LinkedIn, long articles were 50 to 66% of cited LinkedIn content. About 95% of cited posts were original, not reshares (Semrush, vendor). The pattern: other people, writing specifically, in their own names. 2. Collect honest reviews, because small edges beat big names With identical specs, the well-known brand won 100% of the time. A rating edge under 0.1 stars for the smaller competitor flipped it (Chu and Hou, 2026, pre-print). In Columbia's study of shopping agents, +0.1 in rating moved a 10% pick rate to 15 to 20%. A reviewer "pick" badge took it to 20 to 43%. A "Sponsored" tag dropped it to 8 to 9%. A single Wirecutter or Reddit source could override a model's favourite (Wharton GAIL). Social proof raised recommendation rates in controlled tests. Scarcity and "exclusive" cues lowered them (Bias Beware, EMNLP 2025). 3. Show up in the assistant's live search Brands absent from the assistant's web results were mentioned 2.8% of the time. With their own site cited plus branded searches: 91.4% (Tannenbaum, 2026, pre-print). Inside those results, position 1 was cited 58% of the time, position 10 only 14% (AirOps and Kevin Indig). In C-SEO Bench (NeurIPS 2025), simply being first in the model's context beat every rewriting trick. So ranking in ordinary search for "best [category] 2026" is still real work. Not old-fashioned work. 4. Put a clear answer in the first third of the page 44.2% of citations come from the first third of a page. Definite wording gets cited more often (AirOps and Kevin Indig). In the GEO experiments on 10,000 queries, specific numbers, quotations and named sources raised visibility by 30 to 40%. Lower-ranked pages gained the most (Aggarwal et al., KDD 2024). This is the one on-page tactic with experimental support. It boils down to: say plainly what you are, back it up, do it near the top. 5. Show real dates on real updates Models favour newer dates. A more recent publication date moved passages up in all seven rankers tested, by as much as 95 places (Fang et al.). Cited content is fresher than average (Ahrefs). This only works when the date is honest. Bumping the date on an unchanged page is manipulation. If a model ever learns to check, it will cost you. Our rule: a visible date, changed only when the content changes, logged in the changelog (/changelog). What doesn't work llms.txt on its own llms.txt is a plain-text file listing a site's important pages for AI agents. It is a reasonable courtesy and takes an hour. It is not a visibility lever. - SE Ranking found no correlation between having one and AI citations, across about 300,000 domains. Their model got more accurate without it (vendor). - Ahrefs' server logs: 97% of llms.txt files got zero requests in May 2026 (vendor). - Google says it does not use llms.txt for Search or AI features. We have one at /llms.txt (/llms.txt) because agents can use it. We will publish our own server logs on whether anything reads it. Anyone whose deliverable is an llms.txt file is selling a placebo. Schema markup on its own Structured data (JSON-LD) describes what a page is: an article, an organisation, a product with a real price. Use it accurately. No evidence shows it makes an assistant cite or recommend you. Google says no special AI markup exists. Invented ratings in markup are useless and against Google's rules (Search Engine Land's GEO myths review; Google's documentation). Markup that matches the visible page is hygiene, not strategy. Keyword stuffing It did not help in the GEO experiments. Citations, quotations and statistics did (Aggarwal et al., KDD 2024). Most AI-specific content tricks did little in C-SEO Bench, and every trick's gain shrank as more competitors used it. When every brand applied the same optimisation in Chu and Hou's tests, the benefit collapsed from +0.802 to +0.007. Directory blasts Services that submit you to 100 or 500 directories sell volume. ChatGPT and Gemini rarely cite directories. In GetIntel's 9,978 answers they were a minor source next to lists, reviews and community threads. A directory may still give you a backlink for classic SEO. It will not get you recommended. We never sell them. Hidden text, prompt injection, fake reviews These do work in the lab. Hidden text and prompt injection pushed a low-ranked product to the top, including on Perplexity (Kumar and Lakkaraju; Pfrommer et al., EMNLP 2024). That is exactly why they are off the table. They are manipulation, they get found, and they bet the founder's name on a trick that dies at the next model update. Our clean methods (/clean-methods) page lists everything we refuse to do. The honest summary | Tactic | Evidence | Our verdict | |---|---|---| | Named on third-party lists, reviews, threads, videos | Strong, mostly correlational, consistent across studies | Do it. This is the core of our project | | Honest reviews on the sites your assistant cites | Experimental (Chu and Hou; Columbia) | Do it | | Ranking in normal search for your buyer queries | Strong (Tannenbaum; AirOps; C-SEO Bench) | Do it | | Clear answer, numbers and sources near the top of a page | Experimental (Aggarwal et al.) | Do it once, properly | | Real update dates | Experimental on rankers; only honest dates | Do it honestly | | llms.txt | No correlation (SE Ranking); 97% unread (Ahrefs) | Ship it free, expect nothing | | Schema markup | No evidence for citations | Accurate hygiene only | | Keyword stuffing | Did not help (Aggarwal et al.) | No | | Directory blasts | Rarely cited (GetIntel) | No, and we say so | | Hidden text, fake reviews, persona accounts | Works in labs, is manipulation | Never | Limits Most of the "works" evidence is correlational. Much of it comes from companies that sell AI visibility tools (Ahrefs, Muck Rack, Peec, Semrush, SE Ranking). The experiments ran on specific models at specific dates. Effects shrink as tactics spread. None of this lets anyone promise what an assistant will say about you. What to do next Find out which third-party pages your assistants cite for your category. That list is the whole to-do list. The free check (/check) produces it: 5 buyer questions, 3 runs each, 5 assistants. You get who was named instead and the sources cited, within 2 working days. Or run it yourself with the do-it-yourself guide (/guides/check-if-chatgpt-recommends-you). Questions So is GEO snake oil? The on-page tricks mostly are. Getting named on pages assistants already cite, honest reviews and normal search are real, boring and measurable. We only sell the boring part. Should I delete my llms.txt? No need. It costs nothing and some agents may use it. Just do not expect it to change what assistants recommend. Do not pay anyone for one. We have one, and it says the same thing. What about schema markup? Use it accurately where it fits, because it describes your page. No evidence shows it makes an assistant cite or recommend you. Google says AI features need no special markup. If everyone does the things that work, do they stop working? Partly. In one study, when every brand used the same optimisation, the benefit collapsed. Earned mentions and real reviews decay slower than tricks. That is why we prefer them. Sources - Chen et al., University of Toronto, earned media in AI search, 2025 (https://arxiv.org/html/2509.08919) - Muck Rack, Generative Pulse, 207 million citations, October 2026 (https://www.globenewswire.com/news-release/2026/10/28/3390134/0/en/generative-pulse-earned-media-still-drives-more-than-84-of-ai-citations.html) - Ahrefs, best lists research: 43.8% of cited page types (vendor) (https://ahrefs.com/blog/best-lists-research/) - Ahrefs, AI brand visibility correlations, 75,000 brands (vendor) (https://ahrefs.com/blog/ai-brand-visibility-correlations/) - Chu and Hou, brand bias and rating edges, 2026 (pre-print) (https://arxiv.org/abs/2606.17443) - Columbia, What Is Your AI Agent Buying? (https://ar5iv.labs.arxiv.org/html/2508.02630) - Tannenbaum, A fitted stage model of brand mention in AI search, 2026 (pre-print) (https://arxiv.org/abs/2609.23162) - AirOps and Kevin Indig, 815,484 query-page pairs (https://www.getpassionfruit.com/research/is-query-fan-out-a-marketing-buzzword-or-something-relevant) - Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024 (https://arxiv.org/html/2311.09735v2) - C-SEO Bench, NeurIPS 2025 (https://ar5iv.labs.arxiv.org/html/2506.11097) - Fang et al., recency bias in LLM rankers (https://arxiv.org/abs/2509.11353) - Wharton GAIL, agentic shopping technical report, 2026 (https://gail.wharton.upenn.edu/research-and-insights/technical-report-agentic-shopping/) - Peec, most-cited domains by engine (vendor) (https://docs.peec.ai/research/most-cited-domains-by-engine) - SE Ranking, llms.txt across about 300,000 domains (vendor) (https://seranking.com/blog/llms-txt/) - Ahrefs, llms.txt study: 97% of files got zero requests (vendor) (https://ahrefs.com/blog/llmstxt-study/) - Google, AI features and your website (no llms.txt use, no special markup) (https://developers.google.com/search/docs/fundamentals/ai-optimization-guide) - Search Engine Land, GEO myths: this article may contain lies (https://searchengineland.com/geo-myths-lies-467617) - GetIntel, 9,978 AI answers: directories are rarely cited (https://getintel.ai/blog/best-saas-directories-2026/) - Kumar and Lakkaraju, manipulating LLM recommendations with hidden text (https://arxiv.org/pdf/2404.07981) - Pfrommer et al., EMNLP 2024, ranking manipulation for conversational search (https://arxiv.org/abs/2406.03589) - Bias Beware, EMNLP 2025: social proof, scarcity and exclusivity cues (https://arxiv.org/abs/2502.01349) - Semrush, LinkedIn AI visibility study, 89,000 cited URLs (vendor) (https://www.semrush.com/blog/linkedin-ai-visibility-study/)