17 Sep 2026
TL;DR:
Have you been considering AEO or GEO as part of your search strategy? Get in touch with our team directly or via the Contact Us page to discuss how we can help.
AI search is the topic senior marketers and businesses can’t get away from right now. It’s in every board pack, every agency pitch and most LinkedIn feeds, and the message from all three is the same: invest now or watch a competitor get there first.
Conductor’s 2026 CMO Investment Report, a survey of more than 250 enterprise digital leaders, found marketing teams allocating 12% of their digital budgets to AEO and GEO, with 94% planning to increase that in 2026. In other words, AI search is being treated as its own line, sitting beside SEO rather than inside it.
That leaves a lot of marketing leaders in an awkward spot. There’s pressure from above to be seen to act. There’s a market full of AI visibility retainers, many of them expensive and few of them able to show a return.
Our position, after twelve months of working on this with clients, is more measured. AI platforms are changing how people buy and how they enter your funnel, and that deserves serious attention. But most of what’s sold as “AI search optimisation” has little effect on whether you get cited.
What follows is what we’ve found does and doesn’t work, drawn from live campaigns, our own experiments and Google’s published guidance.
Large language models don’t contain a neat database of every company, product and website on the internet. They’re trained on huge amounts of information and learn relationships between words, concepts, brands and entities. That’s why ChatGPT already knows what a company like Salesforce is without needing to search for it every time somebody asks a question.
But when a question is more specific, current, local or commercially focused, things work differently. The AI may take the original prompt and break it down into several smaller searches. This is often referred to as query fan-out.
So a user might ask:
“What’s the best enterprise CRM for a manufacturing company with a large sales team?”
The AI isn’t necessarily searching for that exact sentence. It may search several much simpler queries around:
It then retrieves information from search results and uses those pages to build its response. That distinction matters, because the user’s prompt is not necessarily the search query.
If your website ranks for the underlying searches being used during that process, you’ve got a chance of being included in the information the AI is working from. If you don’t, you’re relying almost entirely on what the model already knows about your brand. That’s why our shorthand at Atomic is
There are two broad ways an AI model can know about your brand.
Parametric memory is information learned during model training. If your brand appeared often enough in the data used to train the model, that knowledge can become part of the model itself. This is slow-moving and tends to favour established brands with strong coverage, mentions and links across authoritative sources.
Retrieval-augmented generation (RAG) is different. The model retrieves current information from the web before answering. This is where query fan-out comes in, and where changes to your website or third-party coverage can influence what gets surfaced much faster.
In practice:
Bottom-of-funnel content can influence retrieval relatively quickly. Brand authority takes much longer to build. You need both.
There is another important distinction too. AI systems don’t retrieve live search results for every question. Research referenced in our work found that retrieval only fired for a minority of tested prompts, with the likelihood increasing significantly as the question moved closer to commercial intent.
That gives us two different jobs. For bottom-of-funnel searches, traditional organic visibility becomes extremely important because the AI is much more likely to go looking for current information. For broader informational questions, brand authority matters more. The model needs to have encountered enough consistent information about your business, expertise and position within a market to understand where you fit.
That second part takes longer. It comes from publishing genuinely useful material, earning coverage, appearing on third-party websites and building a recognisable footprint beyond your own domain.
Google published formal guidance on generative AI search features in June 2026, and despite all the noise around GEO, Google’s message wasn’t particularly revolutionary. AI features still rely on Google’s core search infrastructure. Your content still needs to be crawlable, indexed and eligible to appear in Google Search.
Google’s advice continues to centre around useful, original content, first-hand experience, good technical foundations and established SEO practices. In other words, there isn’t a secret second version of Google that requires a completely different optimisation strategy. There are new behaviours to understand, but the foundations haven’t disappeared.
This is where things start getting messy. A whole list of technical fixes and AI-specific optimisations are now being sold as essential, and the evidence behind many of them is weak.
There’s currently little evidence that adding an LLMs.txt file will materially improve your visibility in generative search. Google has explicitly said it isn’t required, major AI platforms haven’t adopted it as a universal standard and controlled testing hasn’t demonstrated a reliable ranking or citation benefit. It may eventually become useful, but right now we wouldn’t treat it as a meaningful SEO investment.
Structured data still matters. It can help search engines understand pages, qualify content for certain rich results and indirectly improve organic performance. What it doesn’t appear to provide is some direct route into an AI-generated answer. There isn’t a magic piece of JSON-LD that tells ChatGPT or Google’s AI features to cite your business.
Use schema properly because it supports search. Don’t sell it as an AI visibility hack.
Another popular recommendation is to break content into lots of tiny sections because supposedly “AI prefers chunks”. That’s an oversimplification. Clear headings and well-structured answers absolutely help, but deliberately fragmenting an article into dozens of shallow 100 or 150-word sections can make the content worse for the person reading it. AI systems are perfectly capable of understanding context across a full page.
Structure matters. Arbitrary fragmentation doesn’t.
Fake Reddit recommendations, manufactured forum discussions and paid “editorial” mentions haven’t somehow become safe because somebody has relabelled the tactic as GEO. Google’s spam policies still apply. You also have the added problem that AI systems increasingly pull together information from multiple sources, and building an obviously artificial footprint across the web isn’t a particularly sensible long-term strategy.
This is probably the biggest trap. AI has made producing content much faster. It hasn’t made mediocre content more valuable.
We’ve repeatedly seen the same pattern across search: websites dramatically increase publishing volume, visibility rises initially, and then performance falls when Google’s systems determine that the content adds very little original value. High domain authority doesn’t make a site immune either.
The useful question isn’t “how much content can AI help us produce?” It’s “how much genuinely useful content can we now produce without reducing quality?” Those are very different strategies.
Before worrying about AI citations, your SEO fundamentals need to work. That starts with technical health. Google needs to be able to crawl your website, understand your pages, index the right URLs and serve them without unnecessary friction. You don’t need a technically perfect website, but you do need one that isn’t actively making Google’s job difficult.
One issue we see repeatedly is businesses treating indexation purely as a technical problem. Sometimes it is. But often a page isn’t being prioritised because Google has very few signals telling it that the page matters. Internal links, external links, site architecture and overall authority all contribute, and fixing the technical issue without strengthening the page itself doesn’t necessarily solve the underlying problem.
One good page rarely owns an entire subject. Strong organic performance tends to come from covering a topic properly: building a useful group of related pages around the questions, problems, products and buying decisions your audience cares about, then connecting those pages through sensible internal linking.
AI search makes this even more valuable. A single AI prompt can generate multiple underlying searches, and the broader and deeper your useful coverage of a subject is, the more opportunities you have to appear across those searches.
Internal links aren’t exciting, which is probably why they get neglected. But they’re one of the easiest ways of helping Google understand which pages matter, how topics relate to one another and where authority should flow through the website.
Most established websites already have pages attracting links, traffic and authority. The problem is that authority often sits there doing very little. A good internal linking strategy moves some of that value towards commercially important pages that need additional support.
We’ve seen significant improvements from internal linking alone, particularly on larger websites where years of content have been published without a coherent structure behind it. It’s something we look at closely as part of our SEO work.
This is another area where AI search reinforces good SEO rather than replacing it. Pages targeting searches such as:
are commercially critical. They’re valuable in conventional search and they’re exactly the kind of searches AI systems are more likely to retrieve fresh information for.
These pages don’t need to be enormous. They need to be specific, answer the query quickly and give the user enough confidence to take the next step. A 4,000-word landing page doesn’t automatically beat a 1,200-word page. The better page is the one that satisfies the search intent.
This is where we start layering AI-specific thinking on top of the SEO foundations.
Google increasingly talks about unique, non-commodity content, and that’s a useful distinction. Commodity content is information anybody could produce by reading the first five Google results and rewriting them. Original content contains something else: real experience, an opinion, original research, internal data, a case study, a process you’ve used, a result you’ve actually achieved. Something that couldn’t have been written equally well by somebody with no experience of the subject.
AI can absolutely help create content. We use it ourselves, and our own content marketing process can involve a substantial amount of AI-assisted drafting. But the important part is what humans add afterwards. The expertise, judgement, examples, opinions and original information are what stop the finished piece becoming another generic article saying exactly the same thing as everybody else.
Reviews have always mattered. AI makes them harder to ignore.
When somebody asks an AI system to recommend a supplier, software platform, agency, product or service, the system can pull information from review platforms and other third-party sources as part of its research. That means reputation can directly affect whether the system is comfortable recommending you. A business with consistently strong reviews gives the AI evidence that customers generally have a good experience. A business surrounded by negative sentiment creates uncertainty.
Our practical recommendation is straightforward: aim for a strong aggregate rating on the platforms that matter within your industry and build review generation into normal customer processes. Don’t run a frantic three-week review campaign and then ignore it for another two years. Volume, quality and recency all matter.
Video remains badly underused by a lot of B2B companies, and that’s increasingly difficult to justify. YouTube content is highly visible across Google’s platforms and transcripts give search and AI systems a substantial amount of text to understand.
You also don’t need a television production budget. A knowledgeable member of your team talking clearly for 90 seconds about a real customer question can be useful content, and one recording can then be reused across:
The commercial upside isn’t simply “AI likes video”. It’s that video gives your expertise another format, another distribution channel and another opportunity to be discovered.
Links still matter. But when we’re thinking about AI visibility, we’d look beyond links alone, because brand mentions matter too.
Think about what an AI system is trying to work out. Who is this company? What do they do? Are they genuinely associated with this subject? Do other people talk about them? Do independent sources support the claims they’re making about themselves?
If the only website regularly describing your business as an expert in something is your own website, that’s not particularly convincing. Trade publications, news coverage, podcasts, LinkedIn discussions, industry websites, forums, reviews and other third-party mentions help build a much stronger picture. This is why we’d increasingly favour digital PR and brand building alongside traditional link acquisition.
Consistency matters too. If your website describes the business one way, LinkedIn says something different, review profiles use old information and industry directories still reflect an outdated offering, you’ve created unnecessary ambiguity. AI systems are very good at finding those inconsistencies.
This is one of the more practical additions we’ve been making to content strategy. Instead of simply asking “what keyword do we want to rank for?”, we also ask “what searches might an AI system run when somebody asks a commercially important question about this subject?”
Take a target prompt, then investigate the smaller queries generated around it. In some AI platforms you can observe parts of this retrieval process directly, and what you normally see are much shorter, conventional searches. Often things like:
At that point, the work becomes very familiar. We take those searches into Google Search Console and see where the client currently stands. If we’re already getting impressions but sitting somewhere around position 35 or worse, that can indicate the subject deserves a dedicated page. If the query is simply a variation of something we already rank well for, we don’t automatically create another URL. We may instead strengthen the existing page with a dedicated section and a clear answer.
That distinction matters because blindly creating separate pages for every variation can introduce cannibalisation and leave you with several weak pages competing for the same subject.
Run this exercise across the important questions in a customer’s buying process and you start building a content plan around the searches AI systems may use during retrieval. That’s much more useful than bolting “GEO optimisation” onto the end of a conventional keyword report.
There’s a sharper version of AI search happening at the point of purchase. Someone gets two or three quotes, pastes them into ChatGPT, Claude or Gemini and asks which supplier they should choose.
At that point, AI isn’t building a shortlist. It’s comparing named companies and looking for evidence: reviews, case studies, accreditations, trading history and what third parties say about each business.
That makes consistency important. An outdated accreditation, a review profile that has gone quiet or a website claim that doesn’t match what appears elsewhere can weaken the recommendation.
Treat this as an ongoing audit. Keep your credentials, trading history, reviews and strongest case studies accurate and consistent across your website, key review platforms, directories and trade bodies.
AI search is important. It’s changing how people discover companies, research products and compare suppliers, and businesses should absolutely be thinking about it. But that’s very different from saying SEO is dead or that an entirely new discipline has replaced it.
From everything we’ve seen so far, the opposite is closer to the truth. Good technical SEO, authority, useful content, search intent, links, reviews, brand reputation and third-party coverage all still matter. The difference is that AI systems are now taking all of those signals and using them in a different interface.
So yes, there are new things worth doing. Map query fan-outs. Make answers easier to extract. Strengthen your third-party presence. Build recognisable expertise around the subjects you want your company associated with, and think about how AI systems research brands as well as how Google ranks URLs.
But don’t abandon the fundamentals chasing an LLMs.txt file and a dashboard of AI visibility scores nobody can verify. Get the underlying search strategy right first, then build on top of it.
This guide covers the main principles behind the way we’re approaching AI search at Atomic. The full playbook goes deeper into:
