I Classified 90 Tactics From 26 GEO Checklists. Three of Them Weren't Just SEO.

TL;DR: The practical difference between SEO and GEO is about three tactics wide. On August 15, 2026 I fetched 26 GEO checklists and extracted 90 distinct recommended tactics: 70 were ordinary SEO you were already supposed to be doing, 17 were novel-sounding advice with no evidence behind them, and 3 were genuinely different because a language model, not a ranking algorithm, is consuming the page. Only 7 of the 26 articles linked a study a reader could check, and 12 cited nothing at all. GEO is real, but it is a chapter, not a discipline, and it is not a second invoice.

Who is this for?

Anyone being quoted a separate retainer for "GEO" on top of an SEO retainer, and anyone who has to decide this week what to actually do differently. The answer turns out to be short enough to execute in an afternoon.

How did I sort 90 tactics into three buckets?

I collected 27 checklist articles from agencies, SEO tool vendors and AI-visibility vendors, fetched each one on 2026-08-15, and listed every action it told the reader to take. One (thegutenberg.com) turned out to be a gated lead magnet with no checklist on the page, leaving 26. Merging wording variants ("keep paragraphs short" and "one idea per paragraph" are one tactic) left 90 distinct actions.

Then a rule, applied in this order. (a) Ordinary SEO: the action is identical to something documented in Google Search Central or standard practice before generative engines existed. (b) Genuinely different: the action only makes sense because an LLM consumes the page, and I could fetch a primary study supporting it. (c) Unsupported: novel-sounding, with no primary evidence, or with evidence against it.

Evidence tier	The 26 articles
Linked a study you can check (7)	Semrush, HubSpot (both GEO posts), Backlinko, Profound, Position Digital, AI Labs Audit
Named a study but did not link it (7)	Search Engine Land (3 posts), Wellows, LLMrefs, MediaOfficers, AY Automate
Cited nothing at all (12)	201 Creative, Directive Consulting, 321 Web Marketing, PageOptimizer Pro, LLM Pulse, LocalMighty, ClickRank, Stackmatix, Captivate, Amply, GetCito, NoimosAI

What did the count show?

70 of 90 tactics (78%) were ordinary SEO. Crawlability, sitemaps, server-side rendering, Core Web Vitals, HTTPS, schema, alt text, internal linking, author bios, topic clusters, backlinks, Reddit presence, review profiles, Wikipedia, freshness. The "answer capsule" that half the sample sells as a GEO invention is featured snippet optimization, which has been standard since 2016.

17 (19%) were unsupported. llms.txt and llms-full.txt. "Answer Nugget Density of 6+ per 1,000 words." "Five to six citations per page from peer-reviewed journals." "Complete schema equals 47% more Perplexity appearances." "Over 55% of AI queries are voice-based." Sub-1.8-second load times as an AI-specific threshold. Four different mandatory answer-capsule word counts (40-50, 40-60, 40-80, 80-120), none of them sourced. GetCito's checklist tells you to wildcard-allow every bot in robots.txt, which is advice with a security cost.

3 (3%) were genuinely different. They are steps 1, 2 and 5 below.

The citation hygiene is its own finding. HubSpot's GEO best-practices post attributes "40% more AI citations" for visual content to Princeton and Georgia Tech, and links arXiv 2402.14764, which is a paper titled "A Combinatorial Central Limit Theorem for Stratified Randomization." It attributes citation predictors to Arizona State University and links arXiv 2406.01059, which is a paper about image outpainting. Its Content Marketing Institute link for "67% higher AI citation rates" resolves to a hub page containing no such number. That is three broken chains on one page, from the article in my sample with the most citations.

What are the steps?

1. Unblock the AI crawler user-agents, in robots.txt and at the CDN edge. GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot, Google-Extended. This is the only step where inaction is fatal, because on July 1, 2025 Cloudflare changed the default to block AI crawlers for new domains, and on September 15, 2026 it splits search, agent and training into three separate defaults. Failure mode: fixing robots.txt and never checking the edge, where the block usually lives. The robots.txt side of this has its own trapdoors.

2. Make every load-bearing claim survive extraction on its own. Each number carries its own attribution inside the sentence, so a passage lifted without its paragraph still stands up. This is the one content tactic with a real experiment behind it: the Princeton-led GEO paper (KDD 2024) tested nine methods across a 10,000-query benchmark and found quotation addition scored 27.8 and statistics addition 25.9 against a 19.5 baseline on position-adjusted word count. Keyword stuffing scored 17.8, below baseline. Failure mode: converting this into a word-count rule and writing 60-word blocks that say nothing.

3. Do the ordinary SEO, once, and do not pay for it twice. Google's own AI features guidance is unusually blunt: "Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add," and "optimizing for generative AI search is optimizing for the search experience, and thus still SEO." Failure mode: accepting a GEO line item that is your existing technical SEO backlog with a new label.

4. Cut the tactics the evidence contradicts. Ahrefs tracked 1,885 pages that added JSON-LD against 4,000 matched controls between August 2025 and March 2026: AI Overviews moved -4.6%, AI Mode +2.4%, ChatGPT +2.2%, and the authors concluded they "can't tell whether the schema did a tiny bit of good or nothing at all." That is the strongest evidence anyone has, and none of seven AI vendors' crawler docs names schema as an input to answer construction. On llms.txt, SE Ranking's 300,000-domain model found no correlation with citations and improved when the variable was removed, while Ahrefs found 97% of llms.txt files got zero requests across 137,210 domains. Seven of my 26 checklists still recommend it. Failure mode: keeping these because they are cheap, then reporting them upward as work delivered. The full llms.txt autopsy is here.

5. Measure by prompting, and write down that the number is bad. There is no Search Console for ChatGPT. Pick 20 questions your buyers actually ask, run them monthly across ChatGPT, Perplexity and Gemini, and log which URLs get cited. Failure mode: reporting a "Share of Answer" percentage as though it were a rate rather than a sample of a stochastic system.

What actually differs, then?

Three things, and they are real. The click is optional. Pew Research instrumented 900 US adults across 68,879 Google searches and found users clicked a result on 8% of visits with an AI summary versus 15% without. Seer Interactive's 3,119-term, 25.1-million-impression cohort found organic CTR of 0.61% with an AI Overview against 1.62% without. Extraction is passage-level, not page-level, which is why step 2 exists and why page-level rank is the wrong unit. The citation is the product, not the traffic it does or does not send.

That is a genuine strategic shift. It is not 90 tactics wide, and none of it requires a second vendor.

How do you know it is working?

Badly, and you should say so out loud. The least-bad metric is citation count on a fixed prompt set, tracked monthly, reported as a range and never as a trend line from fewer than six months of data. Referral traffic is a floor and a distorted one, for reasons I have already documented in detail: the native GA4 channel misses Perplexity, and roughly 70% of true ChatGPT sessions arrive with no referrer at all.

What would change my mind: a controlled test, on pages not already being cited, showing that any tactic in bucket (c) moves citations. Ahrefs explicitly flagged that their schema sample consisted of pages already receiving 100+ AI Overview citations, so the null result may not generalize to invisible pages. That study is the format the GEO industry owes its customers. Nineteen of my 26 checklists have not run one, and twelve did not cite anyone else's either.