GEO Optimization: A Practical Order of Work

An AI answer about your brand passes through four stages. Each stage has a different owner, and only one of them is mostly yours.

Most published GEO studies test the last stage. They put a page directly into the model's context, rewrite it, and measure whether the model cites it more. The critical survey of 45 GEO studies (arXiv, July 2026) concluded that already-retrieved content can change its own citation, but no technique shows a stable, cross-platform effect on being discovered in the first place.

That changes the order of work. If a crawler can't fetch your page, nothing you do to the wording matters. So GEO optimization starts at the access stage, where the causes are documented by the engines themselves and you can check them mechanically.

Be clear about this split before you plan any work. It keeps budgets honest and stops you paying for promises nobody can keep.

In your control

• Whether your key content is in the raw HTML or only appears after JavaScript runs.
• Which AI bots your robots.txt allows, bot by bot.
• Whether your CDN, firewall or bot management lets those bots through.
• Whether your pages are indexable and eligible for snippets.
• Whether the facts on your own site are specific, current and consistent.
• Where you pursue coverage on other sites.

Not in your control

• What a model learned before its training cutoff.
• Whether an engine runs a web search for a given question. ChatGPT search decides at the model's discretion, and Claude.ai search is a user toggle.
• The queries the engine actually runs. OpenAI says ChatGPT search rewrites a question into one or more targeted queries before sending them to search providers.
• Where your page lands in the model's context, and which retrieved pages it cites.

If someone sells you a guaranteed outcome in the second list, ask what evidence they have. Our answer to whether any proven way to increase AI citations exists covers what the research does and doesn't support.

Step 1: Put load-bearing content in the HTML

Vercel's crawler study found that ChatGPT and Claude crawlers fetch JavaScript files (11.50% and 23.84% of their requests) but don't execute them, so they can't read client-side rendered content. Independent testing in June 2026 found the same for OAI-SearchBot, Claude-SearchBot, PerplexityBot and several others. Googlebot and AppleBot are the exceptions, because they render.

The check: fetch a page with JavaScript off, render it in a headless browser, and compare the text. Anything that only appears after rendering is invisible to most AI crawlers. Prioritise prices, product names, plan details and one-line definitions. The fix is server-side rendering or prerendering for that content.

Step 2: Set robots.txt per bot

OpenAI documents separate bots. GPTBot collects training data. OAI-SearchBot decides whether you can appear in ChatGPT search answers. OpenAI states that each setting is independent, so you can block training and still allow search:

User-agent: GPTBot
Disallow: /

User-agent: OAI-SearchBot
Allow: /

User-agent: PerplexityBot
Allow: /

User-agent: Claude-SearchBot
Allow: /

The most common real error is a blanket AI block that quietly removes ChatGPT search eligibility. Meta is the only operator that ties access directly to citation: it says allowing Meta-WebIndexer helps it cite and link to your content in Meta AI. Note that user-triggered fetchers behave differently. OpenAI says robots.txt may not apply to ChatGPT-User, and Perplexity says Perplexity-User generally ignores it.

Step 3: Test the CDN and firewall

Robots.txt is only half the gate. OpenAI's help centre says eligibility means allowing OAI-SearchBot and confirming your host or CDN allows traffic from its published IP addresses. Bot protection can return a 403 or a challenge page to a crawler your robots.txt welcomes. That failure is silent.

It can also happen without anyone editing anything. Cloudflare announced that from 15 September 2026 it blocks Training and Agent crawler categories by default on ad-displaying pages for new domains and existing free-tier customers. Check your zone settings.

The AI crawler access audit playbook walks through each test. For a quick first pass, the AI Crawler Checker requests a page as each AI crawler and shows which ones your firewall turns away. If you're wondering why this comes before any monitoring, see why to fix crawler access before tracking.

Google states that to be eligible as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Search with a snippet. It also says there are no additional technical requirements and no special optimizations needed. There is no separate GEO path into Google.

Three things to check:

1. Googlebot access. There is no robots.txt token that blocks only AI Overviews or AI Mode. Googlebot rules are the control.
2. Snippet directives. noindex, nosnippet, max-snippet:0 and data-nosnippet keep content out of AI answers on Google, and Bing applies data-nosnippet to its AI summaries. These often arrive by accident, inherited from a template someone used to hide a cookie banner.
3. The Search Console generative AI control. Inclusion is the default, so it's an opt-out. Child properties inherit from the parent, so a multi-property setup can be excluded without anyone noticing. It isn't visible from outside, so look at it directly.

Then check Bing separately. Bing's index backs Microsoft Copilot. In Otterly's 2026 experiment with two AI-generated sites, Google deindexed both and its AI surfaces almost stopped citing them, while Bing kept them and Copilot kept citing them. The same test also showed ranking in Bing produced effectively no ChatGPT citations, so indexing is necessary, not sufficient. The Microsoft Copilot visibility checklist covers the Bing side.

Step 5: State specific facts, once and consistently

Once engines can reach you, give them sentences worth quoting. Stating a specific fact (a price, a number, a one-line definition of what you do) on a page that can be fetched gives the engine something it can quote directly. It is the only on-page signal with positive experimental support, and it promises nothing on its own.

The bigger risk is contradiction. Say you run a dental clinic in Austin. Your pricing page lists one cleaning fee, a blog post from two years ago lists another, and your footer still names an old address. An engine that retrieves the wrong page repeats the wrong fact. Audit your own site for every core fact: prices, plan names, founding year, location, integrations, ownership. Make each one say the same thing everywhere, in the HTML.

Step 6: Be named on the sites engines already cite

For many questions, engines cite review sites, directories, comparison articles and forums instead of brand sites. When that happens, being named accurately on those sites is the route into the answer. The evidence shows a strong association here, not a guaranteed effect. Brand prominence is also the biggest uncontrolled factor in citation studies, so expect large brands to dominate.

Practical moves: correct outdated listings, keep directory profiles current, and earn genuine coverage in the publications your buyers read. Never plant hidden text or instructions aimed at AI crawlers. Research on search agents (arXiv, June 2026) shows injected content can sway some models, which is why it is a misrepresentation risk you should watch for on third-party pages, not a tactic.

Measure each engine on its own. Engines use different indexes and different retrieval rules, and the Otterly experiment showed one engine citing a site thousands of times while four others barely cited it. A blended number would have hidden that. Our page on whether what works on one AI engine works on the others explains why.

Know what the numbers can and can't tell you:

• Retrieval fires per question at the engine's discretion. Any single visibility score averages answers given with and without a live web search.
• Google's Search Console report for AI features shows impressions only, with no clicks and no position. Treat any claimed position inside an AI answer with suspicion.
• Prompt trackers see a sample. Otterly's tracker logged 3,934 citations for one site, while Microsoft's own report credited it with 42,817.

Start with a baseline before you change anything, then retest after each fix. The free checker on this site gives a snapshot of whether AI engines name you for up to three buyer questions. Access fixes can take a crawl cycle or longer to show; see how long AI visibility work takes.

• Allowing GPTBot to get into ChatGPT. GPTBot governs training only. Search runs on OAI-SearchBot.
• Blocking all AI bots in one rule. It removes search eligibility along with training use.
• Treating schema as a quick win. Google says structured data isn't required for generative AI search. Ahrefs' matched test of 1,885 pages found AI Overviews citations fell 4.6% after schema was added. Existing markup does no harm, so there's no need to remove it, but it isn't a lever for AI answers.
• Publishing llms.txt and expecting results. Google says Search ignores it, and no engine documents reading it.
• Rewriting for an AI style or chunking pages. Google rejects both and says there is no ideal page length. C-SEO Bench found most citation-rewriting methods ineffective and often harmful.
• Chasing an E-E-A-T or domain authority score. Google says E-E-A-T isn't a specific ranking factor, and no engine computes domain authority.
• Mass-publishing AI content. Google deindexed both of Otterly's 1,000-post sites with no manual action warning.
• Skipping the firewall test. A perfect robots.txt means nothing behind a 403.