GEO Best Practices: 12 Ranked by Evidence, Not Opinion
This guide ranks 12 generative engine optimization best practices by the strength of evidence behind each one - peer-reviewed, vendor-documented, consistently observed, or folklore. Citevora built the ranking from the GEO benchmark in arXiv 2311.09735, published platform retrieval documentation, and citation data across B2B SaaS accounts. Order matters more than the list: this is a 30-day sequence, not a 12-item audit.
Key takeaways
-
Five practices carry the measurable weight. Citations, quotations, statistics, and primary-source hygiene are the only tactics with a peer-reviewed effect size attached to them.
-
Retrieval access ranks first even though no study measures it. A blocked passage cannot be cited, so a gate outranks a lift.
-
Six practices are cheap consistency work. Answer blocks, tables, entity naming, and cross-profile fact matching cost hours, not quarters.
-
Third-party corroboration is the ceiling. It is the slowest practice on the list and the only one no agency, tool, or schema file can shortcut.
-
Schema and llms.txt are housekeeping. They help machines parse you. They do not earn citations, and any agency leading with them has inverted the order.
Why a ranked list beats a checklist
The US top 10 for generative engine optimization best practices returns nine organic results. Eight are agency blogs. One is a tool-vendor guide. One is a community post. That is the supply side: a lot of people explaining the category, almost nobody telling you what to do on Monday morning.
The top three compounds the problem in three different ways. Straight North lays out its tactics as unnumbered bullets, so priority is whatever order the writer felt like using. Go Fish numbers its practices one through six, which looks like ranking but only encodes sequence - nothing in the framing tells you which of the six moves visibility most. Terakeet gives no ranked list at all and attaches no quantified evidence to its recommendations, so a reader cannot distinguish a tactic that has been measured from a tactic that has been repeated.
That leaves a reader with a generative engine optimization checklist and no way to triage it. Twelve items, one afternoon of capacity, no instruction about which item survives the cut. Undifferentiated advice produces undifferentiated execution: teams ship the easy items (a schema block, an llms.txt file) and skip the hard, high-effect ones (retrieval access, dated statistics, third-party corroboration) because nothing in the list told them those were different in kind.
The commercial data argues the same point from the other side. Across the cluster, search volume sits between 30 and 110 queries a month, while cost per click runs from $13.32 to $28.98. Thin volume with expensive clicks means the people searching are few, senior, and buying. A reader with budget authority does not need another definition of the discipline. They need the order of operations, and the confidence that the order was derived rather than asserted.
How we ranked these
Four evidence tiers, defined narrowly enough that a claim can fail them.
Tier 1 - measured in a peer-reviewed benchmark. A controlled test, published, with a metric that can be re-run. One study in this category qualifies: the GEO paper (Aggarwal et al., KDD 2024), which tested nine content-editing methods across 10,000 queries and reported visibility change on a position-adjusted word count metric. Its limitation is real and worth stating: the engine was a GPT-3.5 simulation fed supplied search-result pages, not a live product test. A peer-reviewed simulation is still the strongest evidence available, but it is directional, not a guarantee.
Tier 2 - documented by a vendor or protocol. The mechanism is written down by an operator or standards body: crawler documentation, robot tokens, retrieval architecture, patents, API references. Documentation proves that a mechanism exists. It rarely includes an effect size.
Tier 3 - observed consistently. The pattern holds across accounts, audits, and re-tests, without a control group. Useful, falsifiable by counter-example, weaker than a benchmark.
Tier 4 - folklore. Repeated widely, no published mechanism, no measurement, no way to fail. Not necessarily wrong. Just unproven, and therefore never first.
Within tiers, we ordered by two rules. First, gates outrank lifts: if failing the practice removes you from the candidate set, it goes higher than a practice that improves your odds once you are in it. Second, effort breaks ties, because a two-hour fix that unblocks a two-quarter strategy should never sit at position nine.
The 12 generative engine optimization best practices, ranked
1. Open retrieval access per agent
Tier 2, documented by platform vendors. Effort: about 2 hours per domain. Time to signal: 2–6 weeks, gated by recrawl.
Every generative answer starts with retrieval, and retrieval starts with a fetch. If your robots.txt disallows the search crawlers, if bot management at the CDN challenges them, or if your answer text only exists after client-side hydration, the passage is not in the candidate set and nothing downstream matters. Check the specific tokens rather than a blanket "allow AI" rule: search-oriented fetchers (OAI-SearchBot, PerplexityBot, ClaudeBot's search path) are distinct from training crawlers (GPTBot), and blocking one while meaning the other is the most common own-goal in the category.
The practical test takes minutes. Pull your ten highest-value URLs with a plain user agent, view source, and search the raw HTML for the sentence you want quoted. If the sentence isn't there, a retrieval fetcher probably cannot see it either. Then check robots.txt, then the WAF rules your security team owns, then rendering. This ranks first despite having no published effect size because it is a gate, not a lift - and because it is the practice most often ceded by default to a security policy nobody in marketing has read.
2. Put a 40–60 word entity-named answer directly under the H1
Tier 3, observed consistently. Effort: about 1 hour per page. Time to signal: 2–4 weeks.
One paragraph, immediately under the heading, that answers the page's question in 40 to 60 words, names the entity, and repeats the target phrase in a way that reads like a sentence rather than a keyword. Long enough to survive passage extraction with enough surrounding context to stay true. Short enough to be lifted whole.
The entity naming is the part most pages miss. "Our approach blends technical audits with content engineering" is unextractable because the subject is a pronoun. "Citevora's how to do generative engine optimization sequence starts with retrieval access, then answer blocks, then statistics" survives extraction intact and carries the brand with it. Watch what the answer block replaces: a 180-word scene-setting intro that begins with a landscape clause. Those intros are the most expensive real estate on the page and they are invisible to retrieval.
3. Add dated statistics with denominators
Tier 1, measured in the GEO benchmark (+21.0% position-adjusted visibility). Effort: 2–4 hours per page. Time to signal: 4–8 weeks.
A statistic without a denominator is an adjective. "Most B2B buyers now start in an AI assistant" is unfalsifiable. "In a Q1 2026 audit of 120 B2B SaaS pages, 41 (34%) were cited by at least one assistant for a commercial prompt" gives a reader, a model, and a fact-checker the same three things: a numerator, a denominator, and a date that makes staleness detectable.
The benchmark result is the strongest single argument for doing the work. Statistics addition produced a 21.0% gain in the study's visibility metric, and the effect compounds with the other two Tier 1 methods. The cost is research time, not writing time. Budget the hours for finding, verifying, and dating the number; the sentence takes two minutes.
4. Add attributed quotations from named practitioners
Tier 1, measured in the GEO benchmark (+28.9%). Effort: 2–4 hours per page. Time to signal: 4–8 weeks.
The highest single-method result in the paper belongs to quotation addition, which is awkward for teams that would rather write than interview. A quotation only counts when it is attributable - name, role, company, date, and enough context that a reader can tell the person actually said it about this subject. "Experts agree" is not a quotation. It is the absence of one with a citation mark after it.
Operationally, this is an interview pipeline, not a copy task. Twenty minutes with a customer, a partner engineer, or your own head of delivery yields two or three usable passages, and those passages get reused across pages, decks, and sales material. Teams that amortize interviews across five pages per month find this practice cheap. Teams that start from a blank page on deadline skip it, which is why it is underrepresented on the pages that need it most.
5. Cite primary sources rather than aggregators
Tier 1, measured in the GEO benchmark (+22.5%). Effort: about 2 hours per page. Time to signal: 4–8 weeks.
Link to the paper, the standard, the filing, the vendor's own documentation - not to the blog post summarizing them. Aggregator chains dilute provenance, and a citation that points at a secondary summary gives a retrieval system one more layer to distrust. This post follows the rule literally: exactly one external link, to the arXiv GEO paper, and no tool vendor or agency blog anywhere in the body.
The discipline shows up in two places. In the prose, claims about mechanism get tied to the document that defines the mechanism. In the markup, sources get rendered as visible links a reader can follow, because an unlinked domain mention is a weaker citation signal than a resolvable URL.
6. Convert comparisons into real HTML tables
Tier 3, observed consistently. Effort: 1–2 hours per page. Time to signal: 2–4 weeks.
Anywhere your page says "compared with", "versus", or "the difference between", the content should be a table with real <table> markup - <th> scope, one row per object being compared, no images of spreadsheets, no CSS-grid divs pretending to be a table. Tables survive passage extraction with their relationships intact, which makes them unusually reproducible in a generated answer, and comparison prompts are where generative engines do the most fan-out.
The failure mode is decorative: a styled pricing grid built from divs, or a comparison rendered as an image because the design team liked the typography. Both are invisible as structure. If you want to see the difference quickly, ask an assistant to summarize your comparison page and check whether the output preserves the rows or collapses into vague prose.
7. Cover the sub-question surface on one URL
Tier 2 mechanism, documented; Tier 3 effect, observed. Effort: 1–2 days per URL. Time to signal: 6–12 weeks.
A single user prompt decomposes into many sub-queries. Each retrieves passages, and the answer synthesizes across them. That fan-out architecture, documented in platform descriptions of how retrieval works, explains why a single comprehensive URL often outperforms a cluster of thin posts: one page can be retrieved for eight sub-queries and attributed each time, while eight short pages each need their own crawl, their own entity resolution, and their own corroboration. The mechanism is documented; the size of the effect is not, which is why this sits below the Tier 1 group.
Practically, this means mapping the sub-question surface before writing - the definition question, the how-it-works question, the comparison question, the cost question, the failure-mode question - and answering all of them under one H1 with clean H2 and H3 scaffolding. Our guide to LLM search optimization covers the retrieval side of this in more depth.
8. Name the entity in every extractable block
Tier 3, observed consistently. Effort: about 1 hour per page. Time to signal: 2–4 weeks.
Every block that can be lifted should stand alone: table cells, FAQ answers, image alt text, captions, section openers. Extraction removes surrounding context, so a block that begins "Our platform does this in three steps" arrives at the model with no subject. "Citevora's ai citation analysis runs three steps" arrives intact.
This is a find-and-replace habit, not a project. Search the page for "we", "our", "it", and "this" at the start of sentences that introduce a fact, and rewrite the ones that carry the claim. The same rule applies inside tables, where header cells frequently say "Us" and "Them" and mean nothing once the table is quoted in isolation.
9. Make entity facts identical across site, docs and profiles
Tier 3, observed consistently. Effort: about 1 day. Time to signal: 4–8 weeks.
Founding year, headcount band, category label, service names, headquarters, and the one-sentence description should match across your site, documentation, review profiles, professional-network pages, and partner directories. Discrepancies do not produce a penalty; they produce ambiguity, and ambiguous entities get resolved against you - a model that finds three different service names for the same company has no reason to prefer yours when describing the category.
Pick canonical strings and propagate them. The one-sentence description is the string that matters most, because it is the one most likely to be quoted. Then check the secondary surfaces: review-site categories, integration-partner pages, conference bios, and press boilerplate written three years ago by someone who has left.
10. Earn third-party corroboration
Tier 3, observed consistently. Effort: ongoing. Time to signal: 3–9 months.
Generative answers reward consensus. When a claim about your company appears in your own pages and nowhere else, it is a vendor assertion; when the same claim appears in independent review profiles, partner documentation, trade press, and customer-published case studies, it is corroborated, and corroborated claims are what get repeated without hedging. This is the ceiling of the discipline, and it is the practice that no amount of on-page work can substitute for.
It is also the slowest, which is why teams skip it and then wonder why citation share plateaus. The work is unglamorous: brief a partner's documentation team so their integration page describes you accurately, ask customers to publish their own results rather than ghostwritten quotes, get into the categories that buyers actually filter on, and give trade reporters something quantified to quote. Start it in month one precisely because it finishes in month six.
11. Ship Article, FAQPage and BreadcrumbList schema
Tier 2, documented as a parsing aid - not a citation lever. Effort: about 2 hours. Time to signal: immediate, marginal.
Structured data tells a machine what a page is, who published it, when it changed, and where it sits in the hierarchy. That helps entity resolution and passage classification, and it can produce rich results where eligibility still exists. It does not make an engine cite you. There is no published evidence that markup moves citation share, and the claim that it does is the single most common unsupported assertion in this category.
The FAQPage case is instructive. FAQ rich results have been restricted to a narrow set of authoritative government and health sites since 2023, so most organizations shipping FAQPage markup are pursuing a search feature they cannot get, on the theory that it improves AI citation. Ship the markup - it is cheap and occasionally load-bearing - but do not sequence a strategy around it.
12. Publish an llms.txt file
Tier 4, folklore. Effort: about 15 minutes. Time to signal: unknown.
A markdown file at the root listing your key pages in plain language, proposed as a way to guide language models toward your best content. It is cheap, harmless, and repeatedly recommended. It also has no published mechanism, no provider confirmation of use, and no measurement. That combination places it at the bottom, and it stays there until someone produces evidence.
Publishing one is fine. Presenting it as a deliverable is not. If a proposal leads with llms.txt, you have learned something about the proposal.
Summary table
|
# |
Practice |
Evidence tier |
Effort |
Impact |
Time to signal |
|---|---|---|---|---|---|
|
1 |
Open retrieval access per agent |
Tier 2 - documented |
~2 hours |
Gate; no access, no citation |
2–6 weeks |
|
2 |
40–60 word entity-named answer block |
Tier 3 - observed |
1 hour/page |
High, compounding |
2–4 weeks |
|
3 |
Dated statistics with denominators |
Tier 1 - measured (+21.0%) |
2–4 hours/page |
High |
4–8 weeks |
|
4 |
Attributed quotations, named practitioners |
Tier 1 - measured (+28.9%) |
2–4 hours/page |
Highest single method |
4–8 weeks |
|
5 |
Primary sources over aggregators |
Tier 1 - measured (+22.5%) |
~2 hours/page |
High |
4–8 weeks |
|
6 |
Comparisons as real HTML tables |
Tier 3 - observed |
1–2 hours |
Moderate, cheap |
2–4 weeks |
|
7 |
Sub-question coverage on one URL |
Tier 2 mechanism, Tier 3 effect |
1–2 days/URL |
High at scale |
6–12 weeks |
|
8 |
Entity named in every extractable block |
Tier 3 - observed |
1 hour/page |
Moderate |
2–4 weeks |
|
9 |
Identical entity facts everywhere |
Tier 3 - observed |
~1 day |
Moderate, defensive |
4–8 weeks |
|
10 |
Third-party corroboration |
Tier 3 - observed |
Ongoing |
Highest ceiling |
3–9 months |
|
11 |
Article/FAQPage/BreadcrumbList schema |
Tier 2 - parsing aid |
~2 hours |
Low |
Immediate |
|
12 |
llms.txt file |
Tier 4 - folklore |
~15 minutes |
Unproven |
Unknown |
Three practices that are housekeeping, not strategy
The bottom of the list deserves its own section, because it is where most agency proposals start. Three items qualify: structural schema (Article and BreadcrumbList), FAQ schema, and the llms.txt file. All three are legitimate. None is a strategy.
Structural schema documents page type, authorship, and hierarchy. Useful for parsing, irrelevant to whether your claim is worth quoting. FAQ schema was once a rich-result play and is now, for most publishers, a machine-readability habit at best. llms.txt is a 15-minute gesture with no published mechanism. Together they total under three hours of work - a rounding error against the 40 to 60 hours that practices 1 through 5 require on a typical commercial site.
That ratio is the diagnostic. A generative engine optimization strategy that opens with schema and llms.txt has inverted the order: it has scheduled the cheapest, least-evidenced work first, usually because it produces a deliverable this week. Ask any partner which of the twelve practices they would cut if you had two hours. If the answer names a Tier 1 method, the ordering is honest. If it names retrieval access, ask why the deliverable exists.
A 30-day implementation order
The sequence below assumes a site with 20 to 100 commercial pages and one person with roughly 60% of their time available. It is deliberately front-loaded with gates.
Week 1 - access and baseline. Day one is measurement, not editing: build a prompt set of 30 to 50 real buyer questions, capture current citation share across ChatGPT, Perplexity, and AI Overviews, and save the raw output as your baseline. Then run the retrieval audit on your twenty highest-value URLs - robots.txt tokens, WAF and bot-management rules, and a raw-HTML check for the sentences you want quoted. Fix what is blocked. This is two hours of assessment and however long your security team takes to approve a rule change.
Week 2 - answer blocks and structure. Write 40–60 word entity-named answer blocks for the top ten commercial pages, then rewrite section openers so every extractable block names the entity. Convert your three most-queried comparison sections into real HTML tables. Ten pages is a realistic week at one hour each plus review.
Week 3 - evidence. This is the expensive week. Add dated statistics with denominators to the ten pages, sourcing each number to a primary document. Conduct two or three interviews and place attributed quotations on the pages where a practitioner's voice carries weight. Replace aggregator citations with primary-source links. Expect four to eight pages, not ten; quality of sourcing matters more than coverage.
Week 4 - coverage, consistency, housekeeping. Extend one or two flagship URLs to cover their full sub-question surface. Standardize the entity description across site, documentation, and profiles. Ship Article, BreadcrumbList, and FAQPage markup, publish llms.txt, and re-run the prompt set to compare against the week-one baseline. Also open the corroboration track in week four - partner documentation, review-site categories, and two customer-published case studies - because it is the longest lead time on the list.
How to know it worked
Measure four things, none of them rankings. Citation share against a fixed prompt set, run weekly, expressed as the percentage of prompts where you appear and the percentage where you are the attributed source. Position-adjusted share, because a mention in the last sentence is worth less than a mention in the first. Retrieval activity, read from server logs - if the search crawlers are fetching more of your URLs after week one, the gate fix worked even before citations move. Referral sessions from assistant domains, treated as a floor rather than a metric, since most assistant-mediated research ends without a click.
That last point deserves emphasis, because it derails more reporting decks than any other. Assistant references and analytics sessions diverge by orders of magnitude: pages with hundreds of assistant mentions can log single-digit sessions, because the user got their answer and left. If your dashboard only counts sessions, your best-performing pages look like failures. The measurement model in how to measure AI visibility covers the full instrument set, and our AI citation analysis service runs the baseline and the weekly re-test for teams without an analyst to spare.
On tooling: you do not need a platform to start. A spreadsheet, a prompt set, and an hourly manual check will carry the first month. Once prompt sets grow past 100 or you need competitor share of voice, the best generative engine optimization tools comparison covers what each platform actually measures - and, more usefully, what it does not.
FAQ
What are the generative engine optimization best practices in order?
Retrieval access first, then a 40–60 word entity-named answer block under the H1, then the three Tier 1 evidence practices - dated statistics, attributed quotations, and primary-source citations - followed by tables, sub-question coverage, entity consistency, corroboration, schema, and llms.txt. The order follows evidence strength and gating: practices that remove you from consideration rank above practices that improve your odds once you are retrievable.
How do you do generative engine optimization on an existing site?
Start with a retrieval audit rather than a content rewrite. Confirm the search crawlers can fetch your key URLs, check whether your answer text exists in raw HTML, and only then edit. Rewriting copy on pages that were never retrievable is the most common wasted quarter in this discipline. From there, work page by page: answer block, then statistics and quotations, then structural fixes.
Does schema markup improve AI citations?
There is no published evidence that it does. Structured data helps machines classify the page, establish authorship, and resolve the entity, which is genuine but indirect value. It is not a citation lever, and the widely repeated claim that Article or FAQPage markup increases citation share has no measurement behind it. Ship it for hygiene, in about two hours, and do not build a strategy on it.
Does llms.txt help with AI visibility?
Nobody has published a mechanism or a measurement showing that it does. It is a cheap, low-risk file that takes about 15 minutes to publish and may eventually be adopted by a platform. Treat it as folklore: harmless to ship, misleading to feature in a proposal, and never a substitute for the five practices that have evidence behind them.
How long before generative engine optimization shows results?
Retrieval fixes can change crawl behavior within two to four weeks; answer blocks and statistics typically surface in four to eight; sub-question coverage takes six to twelve; third-party corroboration runs three to nine months. Citation share is a rolling average, not a switch, so weekly re-tests on a fixed prompt set are more informative than a single audit at the end of a quarter.
Is a generative engine optimization checklist enough?
A checklist tells you what exists. It does not tell you what to do first, and that is the only decision most teams actually face. Twelve unranked items plus one afternoon of capacity produces the easy items, which in this discipline are the least effective ones. Rank by evidence, put the gates at the front, and the generative engine optimization checklist becomes a sequence rather than a wish list.
What are concrete generative engine optimization examples?
Three that travel well: a 50-word answer block under an H1 that names the company and answers the heading's question; a comparison rendered as a real HTML table with scoped headers so rows survive extraction; a statistic with numerator, denominator, and quarter attached, linked to the primary document that produced it. Each is a single block of content, each is auditable, and each maps to a practice above.
How does generative engine optimization work at the retrieval level?
A prompt is decomposed into sub-queries, each retrieves candidate passages, and the passages are reranked and synthesized into an answer with citations. Your page competes at the passage level, not the page level, which is why self-contained blocks that name the entity outperform clever prose that depends on context. The definition guide covers the mechanism in full, and our LLM search optimization playbook covers retrieval behavior per engine.
How do you measure AI visibility without sessions data?
Use citation share against a fixed prompt set, position-adjusted share of voice within answers, and retrieval activity from server logs. Session data undercounts badly, because most assistant-mediated research resolves without a click. Build the prompt set from real sales questions, freeze it, and re-run weekly so the denominator never changes mid-quarter.
What is answer engine optimization (AEO) compared with GEO?
What is answer engine optimization (AEO) answers: the discipline of structuring content so an assistant can lift a direct answer - answer blocks, FAQ format, schema, snippet-shaped prose. GEO adds retrieval-level and corroboration work: fan-out coverage, entity consistency, third-party consensus. In practice AEO is a subset of generative engine optimization best practices, and teams that stop at answer formatting stall at moderate citation share.
The order is the strategy
Twelve practices, four evidence tiers, one sequence. The five that carry measurable weight - retrieval access as a gate, then quotations, primary citations, statistics, and the answer block - deserve the next month of your team's attention. The consistency work in the middle is cheap and worth doing in parallel. Corroboration is the ceiling, and it is the only item that gets harder the longer you delay it.
Everything else is housekeeping. If you are evaluating partners, the useful question is not whether they offer ai search optimization services - most do - but whether they will tell you which of the twelve to skip. Teams that want a ranked plan built against their own prompt set can start with our AI search visibility services overview, read how we sequence engagements on the AI search optimization service page, or see the outcomes in the Citevora testimonial library. B2B SaaS teams usually want the B2B SaaS vertical view; firms selling expertise start with professional services; the full industries index covers the rest.
If you want the 30-day order applied to your site, send us your domain and your five most valuable prompts and we will return a retrieval audit with the tiers marked. That is the same first week described above, and it is enough to tell you whether your bottleneck is access, evidence, or consensus.