Become a source
AI can actually cite.

Here's the thing: GEO isn't a new mystical discipline that replaces everything you already know about being findable online. It's an extension of it. If your site is already thin, unclear, or badly structured, no amount of GEO tinkering will save it. If it's solid, GEO is about making sure AI systems can find, trust, and quote the good parts.

So this isn't a page selling you a shortcut. It's a page explaining the mechanics, the actual research, and what you can do this month that's worth your time.

GEO doesn't replace SEO. It sits on top of it, and overlaps heavily with a related idea called AEO (answer engine optimisation). Same foundation, different output format.

Google's own guidance on AI features says existing SEO best practice applies: your page needs to be indexed and eligible for a search snippet to be a supporting link in AI Overviews or AI Mode. That doesn't mean every AI platform uses the same ranking or selection process. There's no special AI schema, no secret file that unlocks inclusion. You can read Google's guide directly at developers.google.com/search/docs/appearance/ai-features.

Fair enough, you might think, so why does GEO need its own name at all? Because the selection and synthesis step, how an AI system picks and stitches together passages, behaves differently to how a ranking algorithm sorts ten blue links. That difference is worth understanding on its own terms. That's what the rest of this page is for.

Not every AI answer involves a live search. Some come straight from what the model learned during training. Training and live retrieval are different things, and it matters which one is happening when you're thinking about visibility.

When a system does search the live web to answer you, it roughly follows four stages.

1. Query. The system interprets what you're actually asking. It may rewrite your question or split it into several related searches behind the scenes (sometimes called query fan-out), rather than treating it as one single string.
2. Discover. It retrieves a set of candidate sources from the web, using much the same crawling and indexing infrastructure as ordinary search.
3. Select. It chooses specific passages from those candidates, favouring content that answers the question directly and reads as reliable.
4. Answer. It synthesises a response from the selected passages, sometimes showing citations or links back to sources, sometimes not.

This is a simplified model of a search-backed answer, not a published selection formula. Different systems combine retrieval and generation differently. Updating a page can make new information available to search; it doesn't immediately change a model's training.

1. Answer real customer questions, directly

Start with what people genuinely ask, not what you wish they asked. Look at your actual support tickets, sales calls, and reviews for the phrasing customers use. Then answer that question early and plainly, in language a stranger would understand without background context. AI systems are selecting passages that resolve a query cleanly; a page that circles the topic for three paragraphs before answering is a weaker candidate than one that answers in the first two sentences and then elaborates. This isn't about tricking anything. It's about not making the reader, human or machine, hunt for the point.

2. Make your content self-contained

A passage pulled out of context still needs to make sense. Avoid answers that depend on "as mentioned above" or images with no text equivalent. If a paragraph would be confusing lifted onto someone else's screen with no surrounding page, rewrite it. State who the advice is for, what it assumes, and what it doesn't cover, all inside the same block of text.

3. Back claims with real data and explain your method

Original data, explained methodology, and specific figures give a page something to cite. Research from Aggarwal et al. (KDD 2024, arxiv.org/abs/2311.09735) found that adding citations, quotations, and statistics could improve visibility in benchmark tests, with results varying a lot by domain and topic. That's a research finding under test conditions, not a promise about how any specific production AI system behaves today. Treat it as a reasonable, not guaranteed, lever.

4. Show who's behind the words

Make it easy for readers to assess who is responsible for the information. Name a real person or a real organisation. Give them a bio, a role, a reason to be believed. This is a practical credibility recommendation, not proof of a universal AI ranking factor.

5. Keep facts current and mark corrections

When a fact changes (a price, a regulation, a product spec) update the page and date the change. Don't quietly overwrite; note it. Freshness signals matter more for volatile topics than static ones, so judge accordingly. A page about a fixed historical event doesn't need weekly dating. A page about pricing does.

6. Fix the plumbing: structure, crawlability, honesty

Clean HTML text, semantic headings, a canonical tag, a working sitemap, accurate structured data. None of this is exotic; it's the same technical hygiene good SEO has asked for for years. What's not on this list: keyword stuffing, fabricated citations, bought reviews, or paying anyone for guaranteed inclusion. Those aren't grey-area tactics. They mislead readers and can breach search platform policies.

Here's the thing: blocking a crawler and welcoming one look almost identical from the outside. Both involve a robots.txt file, a status code, some quiet server-side logic nobody outside IT ever reads. The difference is in the configuration. Check it before assuming your content is accessible.

Robots.txt is a request, not a lock

Robots.txt tells compliant crawlers what they're allowed to fetch. That's it. It's a note on the door, not a bolt. Any crawler that ignores it, or spoofs its user agent string, walks straight past. So if you're relying on robots.txt as your access control layer, you don't have one.

That matters for firewall and CDN rules too. Don't allowlist traffic based on the user agent string alone; it's trivial to fake. If you need to verify a crawler is genuinely who it claims to be, check the vendor's published IP ranges or verification method instead. And no, a robots rule won't necessarily cover every user-triggered visit; some providers treat those separately from their standard crawl.

Search retrieval and model training are different permissions

OpenAI documents this split clearly: OAI-SearchBot handles search retrieval, GPTBot handles training data collection, and ChatGPT-User covers visits triggered directly by a user's request inside ChatGPT. These can be managed separately. Allowing OAI-SearchBot to surface your pages in search results does not mean you've agreed to GPTBot training on your content. Full detail sits at OpenAI's crawler docs.

Perplexity works differently again. PerplexityBot exists to surface and link sites within answers, not to train a foundation model. Details are at Perplexity's crawler docs.

Two vendors, two different splits. Read the actual docs before assuming your rules do what you think they do.

llms.txt: useful, not magic

llms.txt is a proposed file offering a plain-text summary and navigation aid for AI systems. It's optional. It doesn't replace indexable HTML content, and it doesn't replace your sitemap. Nobody can promise it improves citation or ranking; it isn't an established ranking factor. whatisgeo.app includes one, alongside a full text version kept aligned with the live guide, because it costs little and helps orientation. Fair enough.

Practical checks worth doing this week

Run through these without needing a developer's background, though your developer can help confirm results:

1. Open your main page logged out. Confirm the full answer text loads in the HTML, not hidden behind a script that only fires after interaction.
2. Check the response code. The final intended content page should return 200. Check unexpected redirects and soft 404s.
3. Confirm the canonical tag points where you expect. A wrong canonical can point search engines towards a different preferred page.
4. Use Search Console's URL Inspection tool on your important pages. Site: search isn't comprehensive, so don't trust it as your only signal.
5. Keep your sitemap current. It supports discovery; it doesn't guarantee indexing or inclusion anywhere.

None of this reveals a ranking algorithm, and none of it requires bespoke AI schema markup or guarantees priority placement. It's just the groundwork that determines whether the right systems can see your content at all. At the end of the day, that's the whole job: get out of the crawler's way, check your assumptions, and don't confuse a polite request with a locked door.

Picture a small furniture shop, sells desks, run by two people and a website. A customer types: "Will this desk fit a small home office?"

Generic marketing copy answers that question badly. "Our stylish desk brings functionality to any space" tells the reader nothing they can act on.

Here's a rewrite that does better. (Illustrative values only, not a real product.)

The Elmswood desk measures 120cm wide, 60cm deep, 75cm tall. Measure your available floor space, chair clearance and access before ordering. The top is 120 × 60cm; compare that with your monitor stands and other equipment. Room area alone won't tell you whether it fits.

What changed and why: the answer is self-contained (a passage lifted alone still makes sense), it states a limitation honestly instead of hiding it, and it gives numbers a system can quote directly rather than paraphrase vaguely. None of that guarantees an AI system selects or cites this shop over a competitor's page. But it gives the passage a genuine chance of being useful enough to select, which is the only lever available. Use your real specifications and limitations when applying this to your own pages.

A realistic beginner plan. No agency required, no budget assumed.

Week 1: audit what you already have

• List your top 15 customer questions, pulled from real emails, calls, or reviews.
• Check whether your existing pages answer any of them directly and self-contained.
• Check your important pages with URL Inspection in Google Search Console. A site:yourdomain.com search is a quick clue, not a complete indexing audit.

What to verify: that Google can actually find and index your key pages at all. If it can't, nothing downstream matters yet.

Week 2: fix the foundations

• Add or tidy semantic headings (H1, H2, H3) on your 5 most important pages.
• Set canonical tags where duplicate or near-duplicate pages exist.
• Submit or update your XML sitemap.

What to verify: that each key page has one clear H1, a logical heading structure beneath it, and no canonical conflicts confusing search engines about which version is the "real" one.

Week 3: rewrite for direct answers

• Pick your 5 highest-value questions from week 1.
• Rewrite the top of each relevant page to answer that question in the first two or three sentences.
• Add author or business identity information where it's missing (an about page, a named contact, a real address).

What to verify: read each rewritten answer out loud, alone, with no other page context. Does it still make sense and answer the question fully?

Week 4: add substance and start watching

• Add one genuinely original piece of data, explanation, or detail per page (a real measurement, a real process, a real limitation), with a note on how you know it.
• Pick 3 to 5 prompts real customers might type into an AI tool and run them manually.
• Record what comes back: which platform, which date, what was asked, whether you were mentioned, whether you were cited, and in what context.

What to verify: nothing dramatic yet. You're establishing a baseline, not expecting results in week four. That's normal.

A vendor's visibility score describes its own sample of prompts and platforms. It isn't a universal GEO rank. Start with a repeatable manual baseline, then judge whether a paid tool measures what you need.

The core method: pick a fixed set of prompts that real customers might plausibly ask, and repeat them over time, unchanged. For each one, record:

1. Platform (ChatGPT, Google AI Overview, Perplexity, etc.)
2. Date
3. Exact query text used
4. Whether you were mentioned
5. Whether you were cited or linked
6. The surrounding context (were you framed favourably, neutrally, or not at all)

Do this monthly, or after major site changes, using the same prompts each time so you're comparing like with like.

Two things to keep in mind. First, answers vary. Ask the same question twice and you may get different phrasing, different sources, sometimes a different answer entirely. Some of this is personalisation; the system may weight your location, history, or account differently to someone else's. A single check is a sample, not a universal ranking, and it isn't reproducible for every user everywhere.

Second, separate this from your existing analytics. Search Console reports search impressions, clicks and position, including Google's AI features within its search reporting. That is different from a complete cross-platform citation audit. Track referral traffic from AI platforms in your analytics tool separately when the referring platform passes that information, and treat actual conversions from that traffic as the metric that matters most. A mention with no click and no action is a nice data point, not a business result.