GEO checklist

A GEO checklist is a list of checks that tests whether generative engines, such as ChatGPT, Google's AI Overviews, Perplexity, and Claude, can retrieve a site's pages, use and cite them in answers, and describe the company accurately. This one has 18 checks in six groups. Four groups follow the stages an engine goes through before it cites a page, one covers how changes are tested, and one covers tactics to avoid. A content tactic is included only if research supports it, and each check says how strong that support is. Tactics that did not replicate appear as checks that pass when a plan does not rely on them.

Generative engine optimization (GEO) is work to get a website's pages retrieved, cited, and described accurately in answers written by AI systems. This checklist is the audit companion to the complete guide to generative engine optimization, which explains each stage. Crawler access, indexing, structured data, and the other site-wide checks are in the AEO checklist. Where a check belongs there, this page names it instead of repeating it. The sources behind each check were read on September 27, 2026.

Scope

• Covers: GEO work on a company's own pages for Google's AI Overviews and AI Mode, ChatGPT search, Gemini, Claude, Perplexity, and Microsoft Copilot, and the way changes to those pages are tested.
• Does not cover: crawler access, indexing, structured data, dates, third-party sources, and offer pages, which are in the AEO checklist. The rest of technical SEO is in the technical SEO checklist for AI search, and names and profiles are in the entity and brand consistency checklist. Advertising inside AI answers is not covered.
• For: marketers, content leads, and founders at B2B companies who plan, buy, or audit GEO work.
• Pass criteria are AEO HQ's recommendations. The facts behind them link to their sources.
• Evidence labels:
  • Strong: official platform documentation, or several independent studies that agree.
  • Moderate: laboratory studies, a single study, or correlations.
  • Weak: one small test, or a source with a conflict of interest.
• Verdicts: where a check rests on a research finding, it also gives that finding's verdict in AEO HQ's record of what replicates in AEO and GEO research: replicated, partly replicated, did not replicate, contested, or not yet replicated.

Prerequisites

• The AEO checklist's crawling and indexing checks and its checks on facts in the HTML pass.
• A list of 20 to 50 questions that buyers ask, in their own words.
• Verified properties in Google Search Console and Bing Webmaster Tools. Bing's AI Performance report shows grounding queries, which Bing describes as "the key phrases the AI used when retrieving content that was referenced in AI-generated answers" (opens in a new tab). Grounding means basing an answer on retrieved sources.
• A change log that records which pages changed, what changed, and on what date.
• A prompt log: a way to run the same questions on each engine many times and keep every answer, as described in how to measure AI visibility.

How the checks were chosen

A 2026 review of 45 GEO studies found that "topical relevance and context position are the most reproducible levers, generic heuristics transfer poorly, competition can erode individual gains, and citation-oriented rewrites can impair retrieval" (opens in a new tab) (preprint). It also found that "no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior" (opens in a new tab). AEO HQ built the checklist on that record:

• Groups 1 to 4 check the levers the research supports, with the strength of that support stated for each: being retrieved, relevance to the question, an early answer, and specific, consistent claims.
• Group 5 checks how changes are tested, because no tactic has a proven lasting effect.
• Group 6 covers tactics that did not replicate, are contested, or break platform rules.

Quick checks to run first

If time is short, AEO HQ recommends starting with these five checks:

1. Check 1.1: GEO work starts only on pages that engines already index.
2. Check 2.1: edits add what the question needs, and nothing else.
3. Check 5.1: changed pages are compared with pages that were not changed.
4. Check 5.5: every claimed lift names its conditions.
5. Check 6.1: no page text is rewritten to fit what engines are thought to prefer.

1. Retrieval: whether a page becomes a candidate

Retrieval is the step in which an engine fetches candidate pages from a search index while it builds an answer. A page that is not retrieved cannot be cited, whatever its wording.

• 1.1 GEO work starts only on pages that engines already index. Verify: for each page in the GEO plan, run URL Inspection in Search Console and in Bing Webmaster Tools. Pass: every page scheduled for content changes is indexed in both. Pages that are not indexed go back to the AEO checklist's crawling and indexing checks first. Why: to be shown as a supporting link in AI Overviews or AI Mode, a page "must be indexed and eligible to be shown in Google Search with a snippet" (opens in a new tab), and Bing and Copilot "rely on the same core crawling, indexing, and ranking foundation as traditional search" (opens in a new tab). The 2026 review finds the original GEO paper's gains "valid within its experimental setting but conditional on a source already being present in a fixed context" (opens in a new tab). Evidence: strong (official documentation); replicated (finding 1).
• 1.2 Each priority question is broken into the sub-questions an engine may search. Verify: for each priority question, list the searches an engine is likely to run for it, such as price, scope, comparisons, how the service works, and how results are measured. Add related phrasings from Search Console queries and Bing grounding queries. Pass: each sub-question is answered on a page, or in a section whose heading states it, and no page exists only to catch another phrasing of a question that is already answered. Why: engines often split one question into several searches, which Google calls query fan-out: AI Overviews and AI Mode "may use a 'query fan-out' technique — issuing multiple related searches across subtopics and data sources" (opens in a new tab), and when ChatGPT search uses search partners, it "typically rewrites your query into one or more targeted queries" (opens in a new tab). Google says that creating separate content for every possible variation of how people might search, including fan-out queries, primarily to manipulate rankings or generative AI responses violates its scaled content abuse spam policy (opens in a new tab). We found no study that measures how covering sub-questions changes citations. Evidence: strong for the documentation; the effect is untested.
• 1.3 Rankings are tracked for the sub-questions, in Google and in Bing. Verify: read the ranking report. Pass: it lists each priority question and sub-question, the page meant to answer it, and that page's position in Google and in Bing. Why: in laboratory tests, a page's place among the sources mattered more than its wording. In a NeurIPS 2025 benchmark, making a document the first one in the model's context "leads to far greater citation ranking gains in the LLM response than any C-SEO method" (opens in a new tab), and the authors found that "traditional SEO strategies, those aiming to improve the ranking of the source in the LLM context, are significantly more effective" (opens in a new tab) (peer-reviewed). C-SEO, conversational search engine optimization, is the benchmark's name for rewriting methods. Ranking for the question as typed is not enough: nearly 30% of the domains cited in AI Overviews did not appear in the first page of results at all (opens in a new tab) (preprint; 55,393 queries, March–April 2026). Evidence: moderate; position effects replicated in laboratory settings (finding 2), and AI answers cite beyond the top results (finding 5).
• 1.4 Headings and meta descriptions use the words buyers use. Verify: compare each priority page's section headings and meta description with the questions and sub-questions from check 1.2. The title, H1, URL, and first paragraph are covered by the AEO checklist's check 3.2. Pass: the main terms of each sub-question appear in plain form in a heading or in the meta description, and no common word has been swapped for a rarer or more technical one. Why: in a KDD 2026 laboratory pipeline over 171,003 web documents, extending optimization from body text alone to titles, meta descriptions, headings, and schema fields improved retrieval, "with a +22% boost in Hit Rate" (opens in a new tab), the share of queries for which a page was retrieved. Replacing common words with technical or uncommon ones caused the largest retrieval drops, which the authors attribute to "lexical mismatch between optimized documents and user queries" (opens in a new tab) (peer-reviewed; a keyword-based retriever). In controlled trials across six models, a source that lacked the query's terms was at a disadvantage in which source was cited first (opens in a new tab) (peer-reviewed; laboratory). Google says its AI systems "can understand synonyms" (opens in a new tab), so this check asks for plain words, not every variant. Evidence: moderate; partly replicated (laboratory).

2. Reranking: whether a page reaches the model

Reranking reorders the retrieved pages, usually with a second model, and only the top few reach the model that writes the answer. The AEO checklist's check 3.3 covers the main on-page lever at this stage, an answer at the top of the page. In the KDD 2026 pipeline, placing the answer early "yields higher reranking scores, whereas restructuring that displaces the answer to later paragraphs results in significant rank drops" (opens in a new tab).

• 2.1 Edits add what the question needs, and nothing else. Verify: compare each priority page before and after each edit in the change log. Pass: every added passage answers the page's question or one of its sub-questions, and no edit lengthens a page without adding a fact, widens its topic, or moves the answer lower. Why: in the same pipeline, "the reranker favors content additions that enhance alignment with the query's informational need, while penalizing additions that expand the document's scope beyond what the query seeks" (opens in a new tab). One automated rewriting method, AutoGEO, lost 22.35 retrieval ranks, and the authors trace the drop to "lengthy rewrites that dilute keyword density" (opens in a new tab). 5.8% of target documents dropped from rank 10 to rank 11 during reranking (opens in a new tab), just outside the 10 that reached the model. Evidence: moderate (one peer-reviewed laboratory pipeline); an early answer partly replicates (finding 11).

3. Citation: whether the model uses and cites the page

Once a page is among the sources, topical relevance and list position were the biggest drivers of being cited first in 252,000 controlled trials across six models; explicit price information and a recent timestamp also helped consistently, completeness and trust cues added smaller gains, and formatting-only edits had little impact (opens in a new tab) (peer-reviewed; laboratory; the authors work for Sprinklr, a software company). The AEO checklist covers specific facts in check 3.4 and visible dates in check 4.1.

• 3.1 Claims are stated as plain facts, without hedging or sales language. Verify: read the key claims on each priority page: prices, scope, turnaround, results, and comparisons. Pass: facts the company knows are stated directly, not hedged with words such as "may" or "around"; a fact that is a range is given as the range; and no superlatives or persuasive phrases were added for AI engines. Why: in the six-model trials, hedged language was one of seven factors that made a smaller difference to which source was cited first, along with claims backed by evidence, internal contradictions, missing specifications, and missing comparisons (opens in a new tab). A persuasive or authoritative tone is a different matter: in an ACL 2024 study, models "rely heavily on the relevance of a website to the query, while largely ignoring stylistic features that humans find important such as whether a text contains scientific references or is written with a neutral tone" (opens in a new tab) (peer-reviewed). Evidence: weak for confident wording (one laboratory study); a persuasive tone did not replicate (finding 14).
• 3.2 Claims on different pages agree. Verify: list every page that states a price, plan name, turnaround, feature, or result, including old posts, case studies, and comparison pages, and compare the statements. The company's core facts across the site and its profiles are the AEO checklist's check 5.1. Pass: every statement of the same fact matches, and outdated pages are corrected, redirected, or removed. Why: internal contradictions were among the seven factors that made a smaller difference in the six-model trials (opens in a new tab) (peer-reviewed; laboratory), and Bing asks site owners to "Remove or revise outdated information to prevent incorrect information from surfacing" (opens in a new tab). Evidence: moderate (one laboratory study and platform guidance).
• 3.3 Numbers and quotations support a claim and carry their source. Verify: list every number and quotation on each priority page. Pass: each one answers the page's question or backs one of its claims, and each links its primary source with the sample and date. None was added to meet a quota. Why: claims backed by evidence were among the seven second-tier factors in the six-model trials (opens in a new tab), but numbers added for their own sake did not hold up: in the NeurIPS 2025 benchmark, the statistics-addition method "decreases rankings in 19 out of 24 evaluated settings" (opens in a new tab) (peer-reviewed). Evidence: moderate; adding statistics or quotations as a tactic did not replicate (finding 13).

4. Accuracy: what the answer says about the company

Being cited is not the same as being described correctly. The AEO checklist's check 9.4 grades answers about the company against a fact sheet.

• 4.1 Each wrong answer is traced to its source, and the fix is logged with a date. Verify: for each error found in check 9.4, open the pages the answer cited. Pass: the log records the error, the cited page, whether that page is the company's or a third party's, the correction made or requested, and its date, so later runs show whether the answer changed. Why: 11.0% of 98,020 claims in AI Overviews were unsupported by the pages they cited, with omission the dominant failure mode (opens in a new tab) (preprint; March–April 2026). In a study of ChatGPT, Claude, Grok, and DeepSeek, responses were "largely grounded in search results," but "some claims rely on uncited search results" (opens in a new tab) (preprint), so the cited page is not always the source of an error. We found no study that tests whether correcting a source changes the answers built on it. Evidence: strong that unsupported claims are common (replicated, finding 6); untested that fixes work.

5. Testing changes

No tactic has a proven lasting effect, so each change is a test. The only controlled field study AEO HQ found shows why a comparison group matters: total ChatGPT referrals grew 5.7 times while untreated pages on the same site grew 3.5 times over the same window (opens in a new tab), and the authors conclude that headline AEO multiples "substantially overstate causal effect" (opens in a new tab) (preprint; one site; the authors work for Glasp, the company that owns it).

• 5.1 Changed pages are compared with similar pages that were not changed. Verify: read the test plan for each GEO change. Pass: before the change ships, a comparable set of pages is left unchanged, and both sets are measured over the same weeks with the same prompts and reports. Why: the field study separated the change from the engines' own growth with an untreated part of the same site, and its authors say that "separating treatment from platform tailwind with an on-domain control, matters more than any single multiple" (opens in a new tab). The 2026 review's measurement protocol also includes controls (opens in a new tab). Evidence: strong for the method.
• 5.2 Each test changes one thing, or is reported as a bundle. Verify: read the change log for the test period. Pass: either one kind of change was made, or the report lists every change in the bundle and credits the result to the whole bundle. Why: the field study's estimate comes from a bundle of four changes: URL canonicalization, new pages based on AI crawlers' requests for missing URLs, question-form titles with standalone summaries, and a rule that kept pages with Google clicks from being rewritten (opens in a new tab), so its effect cannot be credited to any one of them. Evidence: strong for the method (a design rule).
• 5.3 Each question is tested in three to five wordings, on several dates. Verify: read the prompt panel and the run schedule. Pass: each priority question has three to five paraphrases, each is run several times per session (the AEO checklist's check 9.1), and sessions repeat on several dates. Why: answers "can vary across runs, prompts, and time, making one-off observations unreliable" (opens in a new tab) (preprint), and the 2026 review suggests "three to five paraphrases per information need" (opens in a new tab) and "Seven to eight repetitions" (opens in a new tab) per prompt as a starting point. Evidence: strong that answers vary (replicated, finding 4); moderate for the numbers (one review).
• 5.4 Each run records whether the engine searched and what it cited. Verify: open ten runs in the prompt log. Pass: each run stores the engine, product, mode, and date, whether a web search ran, the cited URLs, and the full answer (the AEO checklist's check 9.3 covers the product, mode, and account state). Runs with no search or no citation are kept and counted. Why: web-search decisions "vary substantially across platforms and models" (opens in a new tab) (preprint), and in one clickstream panel ChatGPT enabled web search on 34.5% of queries in February 2026, down from 46% in late 2024 (opens in a new tab) (vendor study). The 2026 review asks studies to record whether search was used and to count answers without a search or a citation as results rather than drop them (opens in a new tab). Evidence: moderate.
• 5.5 Every claimed lift names its stage, measure, engine, dates, and comparison group. Verify: read internal reports, vendor proposals, and case studies that report a percentage or a multiple. Pass: each figure states the stage it measures (retrieval, citation, share of the answer, or referrals), the engine and model, the dates, the sample, and the comparison group. A figure missing any of these is treated as an anecdote. Why: different measures give different answers. The 2026 review says the original GEO paper's gains "establish neither organic discoverability nor durable traffic effects" (opens in a new tab), and a benchmark that measured citation rank reported: "Out of 54 cases, we uncover only three where the ranking improvements are statistically significant" (opens in a new tab) (peer-reviewed). Google lists services "Promising improvements for AI experiences and search formats (also known as 'AEO' or 'GEO' tools)" (opens in a new tab) among those to think critically about, and says good advice "either qualifies their claims as opinion based on data or experience, or backs up their claims by citing official Google Search guidance" (opens in a new tab). Evidence: strong.

6. Tactics that did not replicate or break platform rules

Each check in this group passes when the practice is absent. Generative engine optimization antipatterns describes each one with a detection test. Other practices to avoid are covered in the AEO checklist: hidden text, pages built for every variation of a query, false dates, and promises of guaranteed placement in its practices to avoid, and schema markup or llms.txt treated as ways to earn citations in its structured data checks.

• 6.1 No page text is rewritten to fit what engines are thought to prefer. Verify: list the rewriting tools and projects in use, and sample 20 pages changed in the last year. Pass: no page was rewritten by a tool or a prompt "for AI" unless an editor tied each change to a reader's question. Human review of every page is the AEO checklist's check 3.6. Why: in the KDD 2026 pipeline, "optimizing body text alone consistently degrades visibility across all stages" (opens in a new tab) (peer-reviewed). In the NeurIPS 2025 benchmark, "most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking" (opens in a new tab), and the overall gains fell as more competitors adopted the same method (opens in a new tab) (peer-reviewed). Evidence: strong (two peer-reviewed benchmarks agree, as does the 2026 review); did not replicate (finding 15).
• 6.2 Formatting-only changes are not reported as GEO progress. Verify: read the project plan and its reports. Pass: each formatting change, such as bullets, tables, question-and-answer blocks, or splitting pages into small chunks, is tied to a reader's task, and none is reported as a way to earn citations. Why: in the six-model trials, "Formatting choices (Content Structure, Scattered Information) had no impact, suggesting LLMs parse content regardless of visual organization" (opens in a new tab), and Google says "There's no requirement to break your content into tiny pieces for AI to better understand it" (opens in a new tab). Structure does help at retrieval, which check 1.4 covers. Evidence: moderate; contested (finding 16).
• 6.3 No text is written to manipulate a model, whether visible or hidden. Verify: search page text, attributes, and structured data for strings that are not words, product codes, or identifiers, and for phrases addressed to AI systems. Pass: none are found. Hidden instructions are also covered by the AEO checklist's check 10.1. Why: manipulation works in tests: a single polluted page among the retrieved results fooled recommenders up to 27% of the time across 12 models (opens in a new tab) (accepted to Findings of EMNLP 2026). Results depend on the model, from 0.0% attack success on Claude Sonnet 4.6 to 31.4% on Gemini 3 Flash (opens in a new tab) (preprint), and defenses exist: one cut average attack success from 50.32% to 6.20% (opens in a new tab) (preprint). Google's spam policies cover "attempting to manipulate generative AI responses in Google Search" (opens in a new tab). Evidence: strong; manipulation works in tests (replicated, finding 7) and breaks platform rules.
• 6.4 Claims about competitors are dated and sourced. Verify: review every page the company publishes or commissions that names a competitor. Pass: each claim about a competitor carries a date and links a public source, and nothing about a competitor was placed on a third-party site without disclosure. Why: in attacks demonstrated on production search engines (Bing and Perplexity), crafted content could "trick an LLM to promote the attacker products and discredit competitors," and when several parties attacked, the result was a prisoner's dilemma that "collectively degrades the LLM's outputs for everyone" (opens in a new tab) (peer-reviewed). In the polluted-page benchmark, vulnerability increased "when models lack stable prior knowledge of the products" (opens in a new tab). Evidence: moderate (laboratory tests and tests on live engines).

Version and change log

• Version 1.0, September 28, 2026: First published. Sources checked on September 27, 2026.

Next steps

AEO HQ sells generative engine optimization services at fixed, published prices: an automated audit for $499, an audit of search and AI visibility for $2,500, the AEO Blueprint for $4,995, and the Blueprint + Implementation for $8,995.