You Don’t Need a GEO Tool. You Need These Six Things.

Last month, there was a panel in Thessaloniki, “Search Engine Optimization in the AI Era”, at the Alexandria Innovation Zone. I was meant to be on it and couldn’t make it. So this is the page of points I had prepared and never got to deliver, written for the people who need them most and can afford them least: small businesses and one-person marketing teams.

For context: I work at Baresquare, a team that spends much of its week on AI-visibility monitoring for large enterprise clients. So watching these systems contradict themselves is close to a day job.

I asked Gemini and ChatGPT the same question, “Top 3 places to buy running shoes in central London,” and I asked it ten times. Gemini gave me three different number-one answers across six runs. Of the seven shops it named at one point or another, exactly one showed up in all six. The shop in third place was almost a different shop every run. ChatGPT was calmer: the same three shops in all four runs, the same winner every time, with only the second and third spots swapping once. One prompt, ten runs. Not a study, a receipt. And this is the surface: an entire industry now sells you weekly “AI visibility” dashboards to track.

I ran a second test with a different set of prompts, and it came out even less consistent. Only 2 results appeared on more than 1 run when I asked for the “Top 3 GA4 Agencies”.

None of this is a glitch. It is what you should expect from a system that builds each answer fresh against an index, instead of handing back a stored list of links. Ask twice, get two answers.

1. What actually changed

For twenty years, SEO meant one thing: rank high in Google’s organic results. The rules were roughly shared across engines and roughly predictable. Same query, same ten blue links, more or less, every time.

That is over. Search has fragmented from one machine you optimize for into a dozen AI engines that each behave differently and change weekly. Google still holds around 90% of traditional search, but traffic is leaking into ChatGPT, Gemini, Perplexity and the rest, and each one answers in its own way.

The bigger shift is that people stopped clicking. Zero-click searches, where you get your answer without visiting a single site, now sit around 70% of all queries, up from roughly 45% a decade ago (a16z). SparkToro’s 2026 clickstream analysis puts it at 68% for the first four months of the year. And when a Google result carries an AI summary, people click a link inside it about 1% of the time. You can be number one, and nobody sees you.

Here is the honest nuance the doom headlines skip. Only 0.34% of searches actually reached Google’s full AI Mode in early 2026. It is AI Overviews, the summary box sitting on top of normal results, that drives the click collapse today, not AI Mode. SISTRIX measured it: on searches where an AI Overview shows up, clicks on the first organic result fall from 27% to 11%, a loss of almost 60%. AI Mode is the wave still forming. Google says it passed a billion monthly users, with queries doubling every quarter, though that is Google’s own figure, so read it as direction, not gospel.

Then, in May, Google published its first official guide to optimizing for AI search and quietly torched half the GEO cottage industry. The message was blunt.

SEO is still the foundation.

AI Overviews and AI Mode run on the same index and ranking systems, so a page that can’t rank can’t be cited. Write non-commodity, people-first content. And you do not need llms.txt files, content “chunking,” or special markup, because Google’s systems do not use them. The thing that a good share of GEO consultants are selling does nothing.

Although Google can’t quite keep its own story straight. The same season, its Chrome Lighthouse tool added an “agentic browsing” audit that checks whether your site has an llms.txt file. One part of Google says ignore it; another ships a test for it. John Mueller clarified that the check is about agent readiness, not search rankings, which is true and won’t end anyone’s confusion.

There is a mechanism behind the volatility. Traditional search retrieves existing pages and ranks them into a stable list. AI search retrieves and then generates, and it increasingly personalizes: Google’s AI Mode can now read your Gmail and Photos, so two people typing the identical query get different sources. Ranking is repeatable. Generation is not.

And the metric everyone chases, the citation, is mostly vanity. Lily Ray’s June study found that when a brand publishes its own “best [category]” listicle ranking itself first, Google’s AI cites that page but leaves the brand out of the actual recommendation about 69% of the time, often recommending the competitors named inside it instead. Shopify has published dozens of these self-ranking lists; the AI happily cites them and then recommends someone else. Being a “source” can be worth nothing, and often it is a vote for your rival.

2. What a small business should actually do

Here is what I would do if I ran a small business tomorrow. Six things. Not one of them is a subscription.

One: don’t buy a weekly AI-visibility dashboard.

The numbers move too much to mean anything week to week. Brand visibility swings about 30% from one answer to the next, and holds for only about 20% across five identical runs. The overlap between who ranks and who gets cited fell from 76% to 38% in under a year. Most tools fire 250 to 500 queries, report no sample size or confidence interval, and sell you the weekly wobble as “change.” If you track at all, track monthly, and watch the direction, not the absolute score.

Two: use the free measurement stack that now exists.

It has three layers, and no single one sees everything.

• Google Search Console now has Generative AI performance reports showing your impressions inside AI features.
• GA4 added an AI Assistant channel that catches referrals from ChatGPT, Gemini and Claude, though it is forward-only and misses traffic with no referrer, plus Google’s own AI features.
• Your server logs show what the AI bots actually crawled.

Combine the three; don’t pay for a fourth.

Three: fix your technical hygiene.

This is the one checklist worth printing:

Check	Why it matters	Do this
robots.txt, per AI agent	You can be invisible in one engine because of a single line	Review access for GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot, Google-Extended
Content in HTML, not JS-only	AI bots don’t run JavaScript, so JS-only content is invisible to them	Server-render anything you want cited
Sitemap	Helps bots find pages that sit buried	Keep it current and submitted
Schema, high-value only	Helps machines read your entities	Organization, Article, Product, LocalBusiness. Skip FAQ, its rich result is dead
Structure	AI lifts clean, self-contained answers	Clear headings, short summaries, and tables

You can be invisible in ChatGPT because of one line in a file most owners have never opened.

Four: know the 37-day clock.

The first public benchmark on how long a new page takes to get cited put the median at 6.81 days and 90% within 37. So there is a rule of thumb now. If a page is well past that mark, say 40 days, with nothing, that is a setup problem, not a patience problem. A robots block, maybe, or a page buried so deep nothing reaches it. Your logs will tell you which.

Five: write from experience, not volume.

AI surfaces quote clear, first-hand, declarative content and skip vague thought-leadership. Using AI to help you write is fine. Google is explicit that it penalizes commodity content, not AI-assisted content, whoever typed it. What actually moves you into recommendations is third-party proof: other people mentioning you, and real authority somewhere specific. Not calling yourself the best on your own blog, which we just saw, backfires.

Six: be honest about paid.

Google is testing Highlighted Answers, ads placed inside the AI answer and designed to be nearly indistinguishable from the organic recommendations around them. With organic visibility this volatile, if you need a predictable result this quarter, budget for ads. Keep your organic house in order as the base, but stop measuring its return in a visibility score that shakes more than your traffic does.

3. What’s coming

In a year or two, the searching itself gets done by bots. You will tell an agent “reorder my running shoes” and it will query the retailers, compare, and pick, and you will never see a results page. Google’s 2026 announcements point straight at this: background agents that monitor the web for you, agentic booking, even the assistant phoning businesses on your behalf. You will be optimizing for software, not for a person, and the whole funnel gets redrawn.

Commerce then splits into machine-readable or ignored. Google is pushing the Universal Commerce Protocol and a Gemini-powered Universal Cart; OpenAI has its own Agentic Commerce Protocol and Instant Checkout with Stripe. Which standard wins won’t come down to which is best. It will be whichever the biggest players back. Blu-ray didn’t beat HD-DVD on merit; it won because a PlayStation shipped with a Blu-ray drive in every box. So keep a clean product feed and stay current with the standards, or the agent skips you for the next seller.

The tools change character too, from assistants into something closer to employees. The next wave of SEO software crawls, validates, finds gaps, and monitors on its own, daily, without a human in the loop. The AI-SEO tools market is already worth around $2.4 billion in 2026, by one estimate. The grunt work gets automated. The judgment stays human.

Here is the part that should worry a small business. AI search leans toward brands it already knows, both from personalization and from how often they show up in training data. Left alone, it widens the gap between big and small. Your counter is the niche-authority play from the last section: be the obvious, well-referenced answer in one specific place, where an established generalist is not.

The one thing to keep

The best framing I have for all of this is simple: success with AI is not about the tools; it is a mindset shift. You reimagine the workflow; you don’t just bolt AI onto the one you already have.

So here is the shift I would make first. Stop measuring whether AI cites you. Start measuring whether it recommends you. Only the second one pays.