Why GEO Matters Now

Start By Distrusting The Numbers

Almost every dramatic figure about AI search traffic was published by a firm that sells AI search services. That does not make them false. It does mean the burden of proof sits with the reader, and nothing obliges the reader to pick it up.

Three categories of claim circulate, and they deserve very different treatment:

Kind of claim	Example	How much weight it carries
Company-reported usage	OpenAI reporting roughly 800 million weekly ChatGPT users in 2025	Self-reported and unaudited, but the company has legal exposure for misstating it. Reasonable as an order of magnitude.
Vendor traffic study	A 527% rise in AI-referred sessions across 19 analytics properties, 17,076 to 107,100, comparing January to May 2025 against the same months of 2024, per Previsible	Directional only. The sample is the vendor's own client base, the baseline was near zero, and a large percentage from a small base is easy to produce.
Analyst forecast	Gartner's published press release of 19 February 2024 predicts search engine volume down 25% by 2026. The figure you will see quoted far more often, organic traffic down 50% or more by 2028, traces to no page on gartner.com that I could find	A forecast, not a measurement. Useful for framing a conversation with a board. Not evidence of anything that has happened.

The honest version of the growth argument is narrower and stronger: AI-referred traffic is small, it is growing quickly, and it converts at a rate that makes it worth attention well before it is worth reorganising around. That is a defensible sentence. Most of what you will read is not.

The Real Argument Is About Timing, Not Volume

Set the traffic numbers aside entirely. The case for starting now survives without them, because of how the two layers behind an AI answer actually move.

Retrieval moves in days

The layer that reads the live web responds quickly. Publish a genuinely better answer, make sure the crawlers can reach it, and it can appear in cited answers within days or weeks. This is the layer most GEO advice is about, and it is real.

It is also the layer your competitors can match on the same timescale. Nothing you do here is durable on its own. A better page is a good quarter, not a moat.

Training moves in years

What a model knows without looking anything up was fixed when it was trained. It reflects what the web said about you months or years before the model shipped, and it changes only when a successor model is trained on a web that has since been written.

This is the layer that decides whether an assistant with no internet access still names you, whether it describes your category position correctly, and whether it treats you as an established option or an unfamiliar one. You cannot buy your way into it this quarter, and neither can anyone else.

What Actually Compounds

If the slow layer is the one worth starting, it is worth being specific about what feeds it. It is not your homepage copy. It is what the rest of the web says about you, in places worth training on.

• Being described consistently by other people. Independent articles, documentation, comparison posts and community threads that name you and say the same thing about what you do. Consistency across sources matters more than volume of sources.
• Presence in reference layers. Wikipedia, Wikidata and the structured references that flow from them are disproportionately represented in training corpora and in what assistants retrieve. Notability rules apply and cannot be shortcut, which is exactly why the presence carries weight.
• Original material with your name on it. Data you collected, a benchmark you ran, a method you documented. Anything that must cite you to be discussed at all is the strongest position available, because it makes your name part of the fact.
• A stable identity. One name, one description, one canonical URL, used the same way everywhere. Every variant splits the evidence a system has to work with.

None of this is new to anyone who has done digital PR or entity work. That is a feature. The mechanism changed; the substance largely did not.

What The Delay Actually Costs

Being late to GEO is not a traffic outcome in the first year. It is three specific costs, in this order:

1. A description you did not write. If assistants have thin information about you, they fall back on whatever they do have, which is often a competitor's comparison page or an outdated review. Correcting an established wrong description is far harder than establishing a right one.
2. A default recommendation that is not you. When a category has one obvious answer in the training data, that answer is what gets named when a buyer asks generally rather than specifically. Second place in a generated answer is frequently no place at all.
3. A measurement gap you cannot backfill. If you begin tracking today, your first year of data starts today. If a competitor started six months ago, they can tell whether a change worked and you cannot.

Decide Whether This Is Urgent For You

General arguments do not tell you whether GEO deserves budget this quarter. This does, and it takes an afternoon:

1. Open your analytics and filter referrals for the assistant domains: chatgpt.com, perplexity.ai, claude.ai, gemini.google.com and copilot.microsoft.com. Note both the session count and the trend over the last six months.
2. Compare the conversion rate of that traffic against your organic search average. Whichever way it falls is worth knowing, and there is a reason to expect it higher: the reader arrives having already been advised.
3. Ask five assistants the single most commercially important question in your category, phrased as a buyer would phrase it. Write down who gets named.
4. Ask each of them, in a fresh session, "what is [your brand] and who is it for". Compare the answer to what you would have written. The gap between those two is your actual problem statement.
5. Put the four results on one page. Small but rising traffic with a high conversion rate and a wrong description is the strongest case for starting that exists, and it is your data rather than a vendor's.

If that exercise says the channel is negligible and the assistants describe you accurately, you have earned the right to deprioritise this, which is a genuinely useful answer. The more common outcome is that the description is wrong, and a wrong description is not a traffic problem you can wait out.

Key Takeaways

• Most dramatic AI traffic statistics come from firms selling AI services. Ask what the base was before quoting the percentage.
• The strongest argument for acting now is timing, not volume: retrieval responds in days, training responds in years.
• The fastest work has the shortest half-life. What compounds is being described consistently by other people, in sources worth training on.
• Late arrival costs a description you did not write, a default recommendation that is not you, and a year of missing measurement.
• Five prompts and one analytics filter will tell you whether this is urgent for your business. Run those before reading another industry forecast.