Quick Answer: If your content doesn't rank well organically, doesn't cite its sources, and doesn't answer the question in the first sentence, an AI engine has no reason to select it. That's the whole game, whether you're talking about Google AI Overviews, ChatGPT Search, Perplexity or Copilot.
Most GEO guides you'll read online stop at the basics — rank well, add schema, write FAQs — but they miss the parts that actually decide whether you get cited: how Google's AI picks what to lift in the first place, the crawler-level access issues most sites don't know they have, and the fact that visibility now has to be measured, not assumed.
With that in mind, this article lays out the framework we run through every time we take a client's site through this process. So, let’s dive deeper into a full, step-by-step breakdown, alongside practical examples you can adapt directly to your own workflow.
Understand the mechanics before the tactics
Firstly, It's worth understanding the mechanics before the tactics, because it explains why some of the advice below works. Gemini, AI Mode and AI Overviews all run on the same underlying Google index and reasoning pipeline — they're not three separate systems to optimise for, just three different ways the same output gets displayed. That pipeline runs, roughly, in this order:
Query fan-out
A broad question like "best CRM for small teams" doesn't get treated as one request — Google expands it into dozens of related sub-queries (comparisons, pricing questions, integration questions, industry-specific variants) and retrieves content against all of them.
Corpus building
Each sub-query pulls in candidate passages, which get converted into vector embeddings and scored for how closely they match the combined intent.
Pairwise ranking
Rather than a single static relevance score (the old TF-IDF/keyword-density approach), passages are compared head-to-head against each other, repeatedly, until a ranked set survives — the model asks "which of these two better answers the intent," not "does this page contain the keyword."
Citation selection
This is the part people miss: being retrieved, and even being used to help generate the answer, doesn't guarantee a citation. Google's model checks whether a specific sentence in the generated answer corresponds closely enough to a specific passage on your page to attribute it. Vague, hedgy writing can influence an answer without ever getting your brand named as the source.
The practical implication: optimise at the passage level, not just the page level. A page can rank well overall and still lose every citation battle if its individual paragraphs aren't written as clean, standalone, quotable answers.
1. Master Organic SEO Foundations
AI-driven search doesn't replace traditional SEO — it sits on top of it. If a page isn't crawled, indexed and ranking reasonably well already, no amount of "AI optimisation" saves it.
Rank well organically first
AI SEO, isn't that different to traditional SEO. AI tools pull from, and reason over, pages that are already performing in traditional search. There's no shortcut around this.
Get technical crawlability right — and be specific about which bots.
This is the part most businesses get wrong in 2026. It's not enough to check that robots.txt says "allow all." AI engines now run separate crawlers for training versus live retrieval, and they behave differently:
- Training crawlers (GPTBot, ClaudeBot, Google-Extended, CCBot) collect data to train future models. Blocking these doesn't affect whether you show up in an AI answer today.
- Search/retrieval crawlers (OAI-SearchBot, PerplexityBot, Claude-SearchBot) index your site to power live citations. Block these and you're invisible in that engine's answers, even if you rank well on Google.
We audit this explicitly for clients now rather than assuming a blanket "allow" is doing what people think it's doing. Worth checking your CDN too — Cloudflare and similar platforms can silently block AI bots at the edge even when your robots.txt says allow, so the two layers need to agree. On Chupi's site, part of the technical audit was exactly this: confirming crawlability wasn't accidentally being blocked at the robots.txt or CMP-consent layer, which matters as much for AI bots as it does for Googlebot.
Make sure content renders without JavaScript gymnastics.
If your key facts, pricing, or product details only appear after a user clicks a tab or interacts with the page, most AI crawlers won't see them. Server-side rendering (or at least a clean HTML fallback) matters more here than it ever did for classic SEO.
Don't ignore page performance.
Core Web Vitals aren't just a Google ranking factor anymore — slow or unstable pages measurably lose out on AI citations too, on the reasoning that AI crawlers have limited patience and heavy pages are more likely to be abandoned mid-fetch or return an incomplete render. Largest Contentful Paint, Cumulative Layout Shift and Time to First Byte are all worth keeping healthy for this reason alone, on top of the usual UX and Google Search case for doing it. This is exactly the kind of fix we made on Laurelle's site, disabling fade-in animations in the Impulse theme — a CLS/rendering fix that was originally about Core Web Vitals for classic SEO, but does double duty for AI crawlability.
Implement schema markup.
Organisation, LocalBusiness, Article, Product and FAQ schema all help AI models identify who you are, what you sell, and what facts are attached to your brand. On Laurelle Antique Jewellery's Shopify site, we bulk-imported structured metafields via Matrixify — stone type, product type, product style — as proper list fields rather than free-text descriptions. That's not just a filtering fix; it gives AI systems (and Google's Merchant/Shopping graph) clean, structured facts to cite about each product, instead of having to infer them from prose.
2. Structure Content for AI Extraction
Lead with the answer.
Open sections with a direct, standalone summary — ideally 40 to 60 words — before you go into supporting detail. AI systems preferentially lift self-contained summary passages rather than digging through a paragraph to find the point.
Use exact-question headings.
Format H2s and H3s as the actual question a user would type, followed immediately by the direct answer, then supporting context.
Make content extractable.
Bullet points, numbered steps, comparison tables and FAQ blocks are what AI models parse and summarise most reliably. This is also where we caught a lot of the issues on Bidmark's own site during our recent content rebuild — the older service pages had templated copy and keyword-stuffed intros that buried the actual answer three paragraphs down. Rewriting those pages answer-first, with FAQ sets per vertical (ecommerce, healthcare, fintech, hospitality, SaaS, real estate, B2B), was as much about fixing extractability as it was about readability.
Cover the full query cluster, not just the head term.
AI engines fan a single query out into several related sub-questions before generating an answer (this is sometimes called "query fan-out"). A page that only answers the exact headline question, and ignores the three obvious follow-up questions a reader would have, is easier for a competitor's more complete page to out-cite. This is why we build FAQ sections per topic cluster rather than per page — it's covering the fan-out, not just the keyword.
Keep content dated and current.
Pages with inline, dated statistics and visible "last updated" signals get cited more than static evergreen copy. If a page hasn't been touched in a year, treat that as a flag, not a feature.
3. Build E-E-A-T and Brand Authority
Demonstrate first-hand experience.
Original research, real case studies, named authors and expert quotes are what separate a page an AI model trusts from one it treats as generic filler. The LinkedIn Pulse article Liam wrote in his own voice as Bidmark's Performance Director — with a named byline, a real title, and a real opinion — is a small but genuine E-E-A-T signal that a ghost-written, unattributed blog post doesn't have.
Earn third-party citations — but broaden what counts as one.
Backlinks from reputable industry publications still matter, but AI engines also draw heavily from sources that traditional SEO used to treat as secondary: Reddit threads, Trustpilot and Google reviews, G2-style comparison sites, and "best of" listicles written by third parties. Being mentioned accurately and favourably on those surfaces is now its own workstream, separate from classic digital PR link-building. If nobody outside your own website is talking about you, an AI engine has very little independent evidence to trust your self-description.
Build entity co-occurrence.
When you accurately reference recognised, trusted entities in your space — competitors, well-known tools, established studies, industry bodies — in proximity to your own brand name, you build a signal AI models use as a trust heuristic: a page that talks intelligently about the wider landscape reads as more credible than one that only talks about itself. This is a natural byproduct of writing honest comparison and "how it works" content rather than pure sales copy.
Maintain cross-platform consistency.
Keep brand name, positioning and messaging aligned across your website, LinkedIn, review platforms and directory listings, so an AI model isn't reconciling three different versions of who you are.
4. Publish an llms.txt File
This is the one addition most 2025-era GEO checklists don't have yet, because the standard is still young. llms.txt is a plain-text markdown file placed at your site root (yourdomain.com/llms.txt) that gives AI systems a curated, human-written index of your most important pages — a one-line site description followed by grouped links with brief context, rather than making a model guess your structure from a sitemap.
Adoption is still early — most sites that have published one get essentially no requests for it yet — but the cost of creating one is close to zero, and several major AI providers have said they'll respect it where it exists. Treat it the way you'd treat a well-structured sitemap.xml a few years ago: cheap insurance for a format that's still forming. Worth adding to the roadmap for Bidmark's own site alongside the AEO/GEO service pages, and for ecommerce clients like Laurelle and Chupi it's a natural place to link buying guides, size/fit or care content, and policy pages directly.
5. Track and Measure AI Search Visibility
This is the piece that's easiest to skip and the one that makes everything above accountable. Ranking work you can't measure is a story, not a strategy.
Unlike classic SEO, where Search Console gives you one dashboard, AI visibility is fragmented across surfaces — Google AI Overviews, ChatGPT, Perplexity, Copilot — each with its own (often limited) reporting. Our article explains ways you can track AI SEO, but in summary, that means:
- Checking Bing Webmaster Tools' AI performance reporting, which currently gives more visibility into Copilot citations than most other free sources.
- Manually and repeatedly testing your target questions across engines to see whether, and how, you're cited — including from different locations, since AI answers are geo-personalised the same way search results are.
- Tracking branded query volume as a proxy signal, since AI engines lean on brand recognition as part of what they choose to cite.
We've been building out exactly this kind of tracking for client and Bidmark content recently — pairing the content and technical work above with a repeatable way to check whether it's actually moving the needle in AI answers, rather than assuming it is.
6. Don't Ignore Multimodal Content
As AI systems get better at parsing images, video and audio, visual content is becoming a genuine (if still secondary) ranking input rather than pure decoration. Descriptive alt text, transcribed video content, and image structured data are worth doing properly now rather than retrofitting later — particularly for ecommerce clients like Laurelle and Chupi, where product imagery carries a lot of the information a shopper (or an AI shopping agent) needs.
The realistic split
Worth setting expectations here: classic SEO is still the foundation, not a separate bucket. A reasonable way to think about effort allocation for a business with an existing SEO programme is roughly 70% solid organic SEO, 25% GEO-specific work (the structuring, entity, and measurement layer above), and a small remainder for experimentation. GEO doesn't replace SEO — it's what you build on top of a site that's already earning its rankings the normal way.
Sources referenced
- Search Engine Land — How to Optimize for AI Overviews
- LLMrefs — GEO: The 2026 Guide to AI Search Visibility
- LLM Pulse — GEO: The Complete Guide for 2026
- GoGoChimp — Generative Engine Optimisation: The Definitive 2026 Reference
- Capston AI — Robots.txt for AI Bots
- DataImpulse — Robots.txt & AI Crawlers in 2026
Categories:
General |
