Getting recommended by AI assistants comes down to four things: the assistant must be able to reach your pages, understand who you are, find a passage worth citing, and see other sources that agree with you. That is what generative engine optimization (GEO) means in practice for a B2B SaaS company. None of it guarantees a place in an answer, and anyone who promises one is guessing, but each step is concrete, and most of them also help ordinary search.
This guide is what we do on client sites and on our own. It replaces several older posts of ours on GA4 and GEO; the useful parts are folded in here, and the rest is gone.
What is GEO, and how is it different from SEO?
GEO is the work of making your company show up, accurately, when someone asks ChatGPT, Gemini, Perplexity, Claude or Copilot a question your product answers. SEO aims for a ranked link; GEO aims to be named or cited inside a written answer, often as one of a short list.
The two overlap heavily. Assistants that browse the web lean on search indexes and on the same pages search engines rank, and OpenAI says ChatGPT search also uses third-party search providers. What changes is the unit of value: an assistant lifts a passage, not a page. Pages that state a clear answer, with specifics, get lifted. Pages that warm up for four paragraphs do not.
The stakes for B2B are real. In G2’s 2026 survey of 1,076 B2B decision-makers, 51% said they now start software research in an AI chatbot more often than in Google.
Can AI assistants actually reach your site?
Check this first, because nothing else matters if the answer is no.
For ChatGPT, OpenAI says sites must allow its OAI-SearchBot crawler to appear in ChatGPT search. Blocking GPTBot is a different decision: it only opts you out of model training. Many sites block both by accident, through a copied robots.txt or a bot-protection rule at the CDN, and then wonder why ChatGPT describes a competitor instead. Other assistants have their own crawlers; read your robots.txt line by line and decide on each one deliberately.
Three technical checks we run on every audit:
- robots.txt and firewall rules. Allowed in robots.txt but blocked by a bot-protection setting is still blocked.
- Content in the HTML. Don’t count on a crawler running your JavaScript. If your pricing, features or docs only appear after client-side rendering, a crawler may see an empty shell.
- Speed and stability. Pages that time out or redirect in loops are skipped by crawlers of every kind.
This site is built as static HTML for exactly these reasons. If your marketing site is a heavy single-page app, our web development team can tell you what it would take to fix it without a full rebuild.
Does the assistant describe your company correctly?
Ask it. Open ChatGPT and two or three other assistants and type the questions your buyers type: “best [category] software for [industry]”, “[your product] vs [competitor]”, “is [your product] good for [use case]”, “alternatives to [competitor]”. Record who is named, how you are described, and which sources are cited.
Most companies find one of three problems:
- You are missing from answers where you should appear.
- You are described wrongly: an old positioning, a discontinued product, the wrong price model or the wrong market.
- A competitor owns the comparison, because theirs is the only page that addresses it.
The fix for the second problem is entity consistency. Write one plain sentence that says what the company does, for whom and where, and use it everywhere: your homepage, About page, LinkedIn, review-site profiles, partner directories, app marketplaces and press boilerplate. When every source agrees, an assistant has little reason to improvise. When they disagree, it averages them, and the average is rarely what you would choose.
Repeat the same set of questions every month and keep the log. That is the most honest competitor benchmark available: answers vary between sessions and users, so look for trends across a fixed question set, not a single screenshot.
What makes a page worth citing?
A page gets cited when one passage answers a question completely and can stand on its own. In practice:
- Answer first. The first two or three sentences answer the page’s question directly.
- Question-shaped headings. An H2 that reads like the buyer’s question, followed by the answer.
- Concrete specifics. Prices or “from” prices, supported integrations, limits, regions, dates. “As of October 2026” is more citable than “recently”.
- One page per question. A page that tries to rank for everything gets cited for nothing.
- Tables for comparisons, and a short FAQ for the questions that don’t need a full section.
- Named, dated, maintained. An updated date that is true, and an author or team that stands behind it.
We built our own ChatGPT Ads page this way, as an example of the pattern. To find the pages on your own site that should be cited and aren’t, our GEO-first content audit walks through the method, and the SEO and organic growth masterclass covers the wider content strategy.
Why do comparison and “vs” pages matter so much?
Because buyers ask assistants to compare, and an assistant can only cite a comparison that someone has written. If the only “[you] vs [competitor]” page on the web is your competitor’s, that page frames the answer.
Write your own, and make them fair. State the criteria, say plainly where the other product is the better choice, and say who your product is not for. An assistant weighing several sources has little use for a page that claims to win on every row, and neither does a buyer. Useful formats for B2B SaaS:
- “[You] vs [competitor]” for your two or three most common head-to-heads.
- “Best [category] for [specific industry or team size]”, if you can be honest about the alternatives.
- “Alternatives to [incumbent]”, when you are the challenger buyers are switching to.
How much do third-party mentions and reviews count?
A lot. An assistant that recommends a vendor is usually corroborating across sources: review sites, comparison articles, community threads, partner directories, podcasts and the press. Your own site tells it what you claim; other sites tell it whether anyone agrees.
The practical work is unglamorous. Keep review-site profiles complete and consistent with your entity sentence. Ask happy customers for reviews at the moments they are happiest. Get listed in the directories of the platforms you integrate with. Contribute genuinely useful answers where your buyers already ask questions. None of this is new; assistants simply make its absence more visible.
Do structured data and llms.txt help?
They help machines read you; they don’t make anyone recommend you.
Structured data (schema.org markup in JSON-LD) states facts in a form software can parse without guessing: Organization with your official profiles listed as sameAs, SoftwareApplication or Product for what you sell, FAQPage and Article on content, BreadcrumbList for structure. It reinforces entity consistency, which is why we add it to every site we build. It is not a ranking switch.
llms.txt is a proposed convention: a plain-text map of your most important pages, at the root of your site, written for language models. We publish one on this site. Treat it as a hint that is cheap to provide, not as something any assistant is obliged to read.
How do you measure AI referrals in GA4?
Separate the clicks you can see from the influence you can only infer.
The clicks. ChatGPT referrals arrive tagged utm_source=chatgpt.com, and other assistants usually pass a recognizable referrer. In GA4, create a custom channel group with an “AI assistants” channel that matches sources such as chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and claude.ai, and place it above Referral so GA4 assigns those sessions to it first. Then carry the source into the CRM with the lead, so AI-referred leads are judged on pipeline like every other channel. Our analytics service and GA4 work cover this setup.
The reads. When an assistant’s crawler fetches your page, GA4 sees nothing, because crawlers don’t run your tag. Your server or CDN logs do: they show which pages bots such as OAI-SearchBot request, and how often.
The influence. Some people read an answer that names you and never click; they search your brand later or type your URL. Watch for these proxies, against a baseline you recorded before the work started:
- branded search impressions and clicks in Search Console;
- direct traffic to deep pages, not just the homepage;
- a “How did you hear about us?” field on demo forms, which in B2B often catches “ChatGPT” before any tool does.
| Lever | What to do | How you will know |
|---|---|---|
| Crawl access | Allow OAI-SearchBot and other search crawlers; check firewall rules; serve content in HTML | Crawler hits in server logs |
| Entity consistency | One company sentence on every profile; Organization markup with sameAs |
Monthly prompt log: descriptions stop drifting |
| Citable pages | Answer-first copy, question headings, specifics, dates | Pages cited in assistant answers; AI channel sessions in GA4 |
| Comparison content | Fair “vs” and “alternatives” pages | You appear in comparison answers |
| Third-party mentions | Reviews, directories, community answers | More cited sources that name you |
| Measurement | AI channel group, source in the CRM, self-reported attribution | AI-referred leads and pipeline in reports |
What doesn’t work?
Promises of guaranteed placement, for a start. Answers differ between users and sessions, OpenAI says placement is never guaranteed, and nobody outside the assistant companies controls the outcome. Be equally wary of pages stuffed with “AI-optimized” phrasing, of spinning out hundreds of thin pages that each target one prompt, and of rank-tracking numbers built on a handful of sampled answers. They measure noise.
Paying for placement is a separate decision. ChatGPT now sells ads below its answers, and our article on what B2B marketers should know about ChatGPT ads covers that channel. Being recommended organically is the more durable asset, because it does not stop when the budget does.
FAQ
How long does GEO take to show results?
Fixing crawl access can show up as soon as crawlers revisit your pages. Entity consistency, new comparison pages and third-party mentions build over months. Track a fixed set of questions monthly rather than expecting a single moment of change.
Should we block GPTBot?
That is a business decision about model training, separate from visibility. OpenAI says blocking GPTBot only opts you out of training, while OAI-SearchBot must be allowed for your pages to appear in ChatGPT search.
Is GEO different for each assistant?
The fundamentals are the same: reachable pages, a consistent entity, citable passages and outside corroboration. Crawlers, search providers and answer formats differ, so check each assistant your buyers use.
Can Goodish audit our AI visibility?
Yes. We check crawl access, how assistants describe you and your competitors, which pages are citable, and how AI referrals are measured, then give you a prioritised list. Start on the ChatGPT Ads page or contact us.
