Ranking in answer engines.

What GEO actually means

Generative Engine Optimization (GEO) is the practice of shaping web content so that large-language-model products (Claude, ChatGPT, Perplexity, Gemini, Copilot) cite your pages when they answer user questions.

It is not SEO with a new coat of paint. It is a parallel discipline that shares some infrastructure with SEO (structured content, clear headings, schema) but optimizes for a different outcome: not a click, but a citation. The user may never visit your site. They will read your answer inside the model, and your brand name will appear underneath.

Treat it as complementary to SEO, not a replacement. For commercial intent queries, Google still matters. For research and exploration queries, answer engines are increasingly where the user ends their journey.

Why citations, not rankings

Answer engines do not expose a linear ranking the way Google does. There is no “position 3”. Instead, the model summarizes a handful of sources and shows them as citations. Winning is being in that handful.

That changes the incentives:

• Quality of claim beats keyword density. Models cite pages that assert specific, verifiable things. Filler content gets stripped.
• Structure matters more than volume. A 900-word page with clean claims and evidence will often out-cite a 3,000-word page with the same information buried.
• Reputation is source-of-truth level, not just domain-authority level. Models weigh “is this source commonly cited by other credible sources” heavily. This is more brittle to game than DA.

The three patterns that get cited

We have tracked citation behavior across 400+ queries (see the share-of-answer data). Three content patterns appear overrepresented in citations:

Pattern 1: the claim-evidence-source structure

Make the claim. Back it with evidence. Attribute the evidence to a source. In that order, and visibly.

Example (bad):

Many experts believe that prompt caching can significantly reduce API costs in production.

Example (good):

Prompt caching reduces input-token cost by roughly 90 percent on cache hits (Anthropic pricing documentation, 2026). In a production workload with a 10,000-token system prompt and 60 percent hit rate, we measured a 48 percent reduction in monthly spend.

The second version names the claim, quantifies it, attributes the source, and provides concrete evidence. That is what gets cited.

Pattern 2: the named framework or taxonomy

Models love lists that have shape. “The five service lines,” “the seven pillars of IAM hardening,” “the three patterns that get cited.” A named taxonomy with a consistent count is memorable, quotable, and cite-able.

Do not manufacture frameworks for the sake of having them. Do name the structure when you actually have one. Make the count explicit in the heading. Make each item internally consistent in shape.

Pattern 3: the first-party experience report

Models distinguish, roughly, between secondary analysis (“here is what the literature says”) and primary experience (“here is what we measured in production”). They cite the second more often when the question is practical.

If you have run the workload, shipped the feature, measured the cost, conducted the pentest - say so explicitly. “In our production deployment we observed X” carries more citation weight than “studies suggest X.”

Mechanical basics: schema and structure

Schema.org is not a ranking signal for answer engines the way it used to be for Google rich results. But it is a clarity signal. A page with clean Article, FAQPage, HowTo, Organization, and Author schema is easier for a crawler to parse into structured claims.

Our minimum schema bar for GEO-optimized pages:

• Article with author (named human), datePublished, dateModified, description
• Organization schema for the publisher with logo, url, and sameAs social profiles
• FAQPage for any content that answers discrete questions
• HowTo for procedural content

Structural rules that matter more than schema:

• One clear claim per paragraph. Multi-claim paragraphs get paraphrased badly.
• Heading hierarchy matches the document shape. H2 for sections, H3 for subsections. Do not skip levels.
• Short, literal heading text. “Why caching matters” is better than “Unlocking the secrets of prompt cache velocity.”
• Tables for comparisons. Models parse tables well and often lift them into citations.
• Explicit definitions. A dedicated paragraph that defines the term you want to own, early, with the definition phrased as “X is Y that does Z.”

llms.txt and robots.txt

Publish an llms.txt at your root. It is an emerging standard that gives LLM crawlers a curated map of your most citable content - often a tighter version of your sitemap with annotations.

Do not block LLM crawlers in robots.txt unless you have decided, deliberately, that you do not want to be cited. Blocking ClaudeBot, GPTBot, PerplexityBot, or GoogleExtended has an immediate, measurable effect on your share-of-answer. Some publishers do choose to block; that is a strategic call with real tradeoffs, not a default.

Authorship and entity signals

Models are notably more likely to cite content tied to a named author with a verifiable track record. Pages with no byline, or with “by Admin”, get downweighted.

The fix is cheap:

• Named author on every page (we sign ours).
• Author page per author, with schema Person, photo, bio, list of work.
• Consistent entity across LinkedIn, GitHub, site, and any other public profiles. Same name, same photo, cross-linked.
• Organization entity disambiguated (full legal name, incorporation jurisdiction, year founded).

Measuring share-of-answer

If you are not measuring, you are guessing. The share-of-answer discipline is simple:

1. List 50 to 400 target queries across your ICP.
2. Run each query against the 4 major engines (Claude, ChatGPT, Perplexity, Gemini) on a fixed cadence - weekly is fine to start.
3. Record whether your domain was cited, what specifically was cited, and how your brand is characterized in the answer.
4. Aggregate into three numbers: citation rate per engine, total citations per query, and sentiment of characterization.

This is what our Share-of-Answer Monitor automates. You can do it manually as well - the discipline matters more than the tooling.

Common GEO mistakes

• Treating GEO as a keyword exercise. There are no keywords. There are claims. Rewrite the content strategy around claims you want to own.
• Hiding evidence in PDFs. LLM crawlers vary in how well they handle PDFs. Put your claims in HTML.
• Walling off content behind a login. Walled content is not citable. If you want attribution, some layer must be public.
• Overloading with schema. Ten schema types on one page confuses parsers. Three or four clean, correct schema types beat ten sloppy ones.
• Neglecting updates. Models weigh freshness. A page with datePublished: 2021 and no dateModified loses citation weight over time. Review and re-date quarterly.

When GEO is not the right play

GEO is high-leverage for informational queries (“what is X,” “how do I Y,” “compare A and B”). It is less useful for:

• Transactional queries where the user needs to convert on your page, not read an answer. Classic SEO plus paid search is still the playbook here.
• Local search (“coffee near me”) where Google and Apple Maps dominate and will for a while.
• Brand-navigation queries where the user already knows you. Do not waste effort GEO-optimizing queries that are already conversions.

When to talk to us

Our SEO & GEO engagements start with a share-of-answer audit against your current content and ICP queries, then produce a 90-day plan plus a content calendar. Start a conversation if you want to know where you stand today.