June 30, 2026
We’ve stayed relatively quiet on AI in public, and that’s been deliberate.
After all, there’s been a lot of noise that we wanted to distance ourselves from. The loudest opinion repeated ad nauseum has been that ‘GEO (Generative Engine Optimisation) now replaces everything you knew about search and that SEO, as well as many other industries, have become obsolete.’
People don’t want to seem like luddites and buzzwords land well in board meetings, so it’s understandable. A serious problem arose, however, when GEO tactics were promoted that actively harm both traditional organic search performance and AI visibility.
Listicle spam and content chunking are prime examples of this, and they’re practices that Google has explicitly stated that it won’t reward. We expect that the other platforms will catch up and eventually implement the same rules.
The reality is more measured: it’s mainly the same playbook, good foundational SEO, with some genuine adaptation around the edges, and a fair amount we (or indeed anyone else) honestly don’t know yet.
After attending the Croatia SEO Summit, where a lot of our own thinking was echoed by people we respect, we’re finally seeing a more grounded, evidence-led outlook on AI become mainstream. So, it feels like the right moment to set out where we stand.
But, none of this means we’ve been sitting on our hands. AI has been baked into our strategy and our day-to-day operations for the last couple of years. We’ve been testing, iterating and improving along the way. Just without making the outlandish claims. With this in mind, here’s what we actually think about it…
First and foremost, the efficiency gains are genuinely phenomenal. AI tools have made us at least four times more efficient than we used to be across research, analysis, production and reporting.
We’re still doing the groundwork to push that further and tighten our processes. For example, we haven’t quite gotten our Claude workflows and skill files perfect yet. However, the point of the efficiency isn’t the efficiency itself - it’s what it frees us up to do.
Faster execution on the mechanical work means more of our time goes into the things that actually move the needle for clients: creative thinking, strategy, and genuinely original ideas. That’s the trade we want, and it’s the one we’re passing on to our clients.
Optimising for AI search beyond doing good SEO is, and remains, far from standardised. There generally aren’t any accepted protocols in the industry yet. There’s no settled answer on whether AI crawlers and LLMs should read an llms.txt file, whether a schema map meaningfully changes how a model interprets a page, or even whether a given model can process structured data at all, which varies model to model.
What people are calling GEO tactics are, for the most part, things a competent SEO strategy already covers. Logical page structure, answering questions succinctly, clean and fast-loading markup, sensible internal linking, strong EEAT signals, great reviews, mentions across the web - none of it is new. These things all assist and convince users in the first instance.
Schema is good SEO - and that’s exactly why it matters for AI, just not in the way that it’s usually sold by some in the industry. The honest mechanism is indirect: strong structured data helps Google understand and index your pages and feeds its Knowledge Graph, and that’s what Google’s own AI features (AI Overviews, Gemini) and any AI agent querying Google directly then draw on.
Google itself calls structured data “critical for modern search features, including generative AI,” while also being clear that no special markup is required to appear in AI generated results. So, schema earns its keep as part of solid SEO, and it’s the SEO that influences the AI - not the markup acting on the model directly.
It’s worth being specific about who actually reads them, because the answer differs by system. The index-based players - Google’s AI Overviews and Gemini, plus Bing and Copilot - genuinely use structured data, because they sit on top of search indexes built with it. The standalone LLMs reading a page live - ChatGPT, Perplexity, Claude - largely don’t: tests show they treat JSON-LD as just more text on the page, and will read fabricated schema as though it were real. So “does schema help AI?” has two answers depending on the system, and the one that matters most today runs straight through Google.
That’s also why most of the tests cited to ‘prove’ that schema drives AI citations run into a correlation-versus-causation problem: the sites with schema in place tended to be the sites taking SEO seriously in the first place. SEO was almost certainly the influencing factor, not the markup itself.
We really liked what Gerry White of Dergal shared with us about the clear benefits of schema for SEO and how it’s a great future-facing thing to do for AI optimisation. AI companies are now under increasing pressure to be profitable, so it’s not hard to imagine a future where these data efficiencies will be leaned on by LLMs.
Markdown conversions, whereby a page's content is made available in a lightweight, LLM-friendly markdown format, absolutely fall within the non-standardised approaches to GEO, with conflicting opinions from the major AI players. As with schema though, it's a great future-facing thing to do.
If you are looking into markdown files for the purposes of GEO, however, Alex Moss of Yoast points out the importance of not duplicating URLs but rather creating different formats of the same URL.
This is the distinction between content negotiation (one URL, format switches based on what the requesting client asks for) and the more common practice of publishing a separate .md copy of every page alongside the original HTML page. The latter is the approach Google and Bing have specifically warned against.
Indeed, if there’s a switch you can toggle through something like Yoast or Cloudflare, by all means switch them on. But, if these providers don’t form part of your techstack, then rather than manually building and maintaining multiple pages, your time will likely be better spent producing content with unique insight, proprietary data, and genuine EEAT.
Grounding pages, or pages built specifically to feed information to LLMs, are one of the more fashionable GEO tactics, and they can work. The logic is sound enough: give a model a clean, structured source and you improve your odds of being drawn on, and drawn on accurately. But, we'd be wary of building them as a standalone exercise.
A page that exists only to feed a machine is usually a page no human ever benefits from, and you're spending effort on a behaviour the models could change at any time. The better return, almost always, is to weave that same information into the content you're already publishing - the pages that serve customers first and happen to be readable by machines second. Done well, you get both: a resource real people actually use, and one the models can pull from, without maintaining a parallel set of pages whose only audience is an algorithm.
Use AI to make yourself more efficient. Use it to spin up scaled content en masse at your peril.
In some cases, mass-produced content will work. However, in the vast majority of instances, it will only work for a very short time. Your traffic may increase over a few weeks, but it will inevitably drop off a cliff and you’ll have a very big and potentially unfixable problem on your hands.
If you’re playing the long game and genuinely trying to build a brand and website that both people and search engines love, the truth hasn’t changed: people are only interested in brands with people behind them who have something unique and original to say - something a reader can connect to, find genuinely useful, or be excited by. AI can help you produce that faster. It can’t be the thing that has the idea.
A handful of on-page habits matter more here than they used to: clear TLDRs and summaries, getting to your point early, descriptive headings that signal what each section covers. None of it is new but it carries extra weight because of how AI systems read. They rarely take in a page the way a person reads top to bottom: content is stripped to clean text, chunked, and retrieved in parts, and even models with huge context windows recall what's near the top and clearly structured far more reliably than detail buried deep. Front-loading your value, saying the important thing early, in plain language, under a heading that flags it, makes it likelier the right passage is the one that gets pulled and quoted, whether you call it SEO or GEO.
It's also worth optimising for query fan-out. Rather than answering a single search, systems like Google's AI Mode break one query into a cluster of related sub-questions, run them in parallel and stitch the results into one answer. But strip away the label and this is fundamentally long-tail keyword targeting with a topic-cluster slant, the same logic behind the pillar pages strategy we've built for years. Cover a topic and its obvious follow-on questions comprehensively, and give each sub-question a clean, self-contained answer the model can lift on its own. The one new caveat is that those answers need to stand up in isolation rather than just contribute to a page that ranks — but the underlying discipline is familiar: topical authority, done properly.
One lever that is genuinely worth highlighting, though, is original, quotable data. Pages built around proprietary stats, survey findings and clear, citable lines tend to get pulled into AI answers far more readily than rehashed commentary or basic who/what/where/why/when/how - which is really just another argument for having something of your own to say rather than repeating what everyone else already has.
Something that Lily Ray of Algorythmic has been very vocal about and that resonates with us is that the web is filling up with low-effort, AI-assembled content produced at a scale - thin listicles and near-identical “best X for Y” pages are the most visible example, but they’re only one flavour of a much wider set of shortcuts being passed off as strategy. It’s the modern version of mistakes we’ve all seen before, and we think a wave of spam updates and manual penalties is on the horizon - and that some surprisingly big brands are going to get caught by it.
It’s easy to forget, but a lot of this was settled in the early days of Google. Content farms, doorway pages and spun articles all worked brilliantly, right up until an algorithm update wiped them out overnight.
The lesson then was the same as it is now: shortcuts that exploit the system get priced in and then penalised. What’s changed is that AI has made the shortcut cheaper and faster, which has tempted some established brands - ones who really should know better - to cast aside best practice in the pursuit of new technology. Chasing scale over substance didn’t end well last time, and we don’t expect it to end well this time either. We’d rather build for clients on the right side of that correction.
If there’s a corner of this where standardisation is racing ahead, it’s agentic commerce - AI agents that research, compare and complete purchases on a shopper’s behalf.
Shopify is at the forefront, co-developing the Universal Commerce Protocol (UCP) with Google, backed by the likes of Visa and Mastercard. There is legitimate, practical preparation to be done here. Whether it ultimately takes off at the scale being predicted is still an open question - but unlike a lot of the GEO chatter, this is real infrastructure being built, and it’s worth watching closely.
That said, I think these models have a long way to go in building up enough trust for you to hand over your card details to them. It’s also difficult to imagine at this moment in time how
Traffic for publishers is not in a good place. As pointed out by Harry Clarkson-Bennett of DJB Strategies, content consumption is at an all-time high, yet younger audiences are only spending 1% of their time on new sites. The industry is fragmenting and AI adds another hurdle to contend with.
From our experience, AI visibility simply does not convert into significant web referral traffic for this industry, and thus ad revenue. So should you open the gates of hell and let the AI crawlers run amok on your website?
The dog has largely bolted. Blocking them now is mostly shutting the stable door after the fact. It rarely sends meaningful referral traffic back to you - these tools are designed to keep the user on the platform, not pass them on.
But, it’s better to look at it as a form of brand PR rather than a channel. After all, someone else is going to be cited (if anyone is cited at all) for your information and ideas. So, you may as well be there and hope the increased discoverability across these platforms pays a return on through direct traffic or other channels. Who knows, there are even rumours of something that hints towards remuneration.
This is the double-edged sword of AI for content creators. On one hand, it creates an incredibly hostile environment that ultimately does not reward people for creating value and things that people enjoy. On the other, it forces a pivot away from commodity content toward more interesting, insightful, proprietary material that AI simply can't replicate - survival of the fittest.
If anything, AI raises the value of digital PR. Mentions, positive sentiment, and being talked about across the web are increasingly what feed both search and AI systems' picture of your brand — arguably the most influential signal there is.
Ahrefs' study of 75,000 brands found web mentions correlate with AI visibility roughly three times more strongly than backlinks, the factor that traditionally sat at the heart of SEO. I'd add the obvious caveat that this is a strong correlation rather than proven causation - brands that get talked about tend to be doing plenty else right but the direction of travel is hard to ignore. Reddit and user-generated content have also become particularly important.
We’re going to have to move away from precise attribution - and we’ve been heading that way for a while. Google has progressively given us less overall and more segmented data from the tools like GSC. Privacy legislation from Apple has compounded it.
The AI tracking tools we’ve tested give an illustrative picture at best. They surface patterns and rough outlines, and we’ve tried several without being especially impressed.
That said, by their very nature, they can’t be more than indicative. We’ve moved away from keywords as a unit of truth, and AI answers are non-deterministic and personalised - ask the same question twice and you’ll get two different answers, and two users almost never see the same response. In that world, referral sessions and a general directional read are close to all we have. The honest shift is towards topical visibility rather than position-level precision.
Keyword tracking tools like Ahrefs and Semrush are nowhere near as reliable as they once were, partly because the SERPs themselves have become so volatile.
We are starting to get more signals - AI Overview impressions are now surfacing in Google Search Console, for instance - but AI search is really just the newest part of a much older trend: people increasingly discover and research your brand on platforms you have no control over and little visibility into. The fact that LLM referral traffic tends to convert well is testament to that - these are people who’ve already done their research before they reach you. It’s the principle that’s always held: non-branded discovery creates branded demand.
So, the reporting will change. Keyword ranking reports are likely to matter less. Average position may become one of the few stable numbers we have. Even traffic and revenue by channel may become more indicative than absolute, as channels increasingly work together in a combined effort to deliver against KPIs rather than each claiming its own slice.
SEO has shifed into a top-level discovery channel that spans multiple platforms rather than one search box. And Google still has a stranglehold on the industry - it remains the primary source of information feeding most other LLMs, and it’s still growing. It holds around 90% of the traditional search market a year on from AI Overviews launching, and even once you fold in the standalone AI platforms like ChatGPT and Perplexity, it stays comfortably dominant at roughly 80%.
Which brings us back to the oldest adage in the business, and the one we’d stake our position on: think about your users first, then make sure your pages can be found and understood through solid technical optimisation. Be on the front foot. Watch which protocols actually become standardised rather than chasing every one that’s proposed. Keep investing in SEO but bear in mind its scope is now much larger than just traditional search.
