Where AI Search is Heading

Why Most Predictions In This Field Are Worthless

The GEO industry runs on forecasts because forecasts sell services. A prediction cannot be wrong yet, which makes it the safest possible thing to publish and the least useful thing to read.

So before the substance, the standard this lesson holds itself to: every direction below is marked as already happening, strong signal or speculation. Act on the first. Prepare for the second. Ignore the third until it moves up a category.

Already Happening

Assistants that act, not just answer

The clearest shift underway is from answering a question to completing a task. Agentic systems already browse, fill forms, compare options across sites and carry a multi-step job to completion rather than handing back a summary.

The consequence is concrete and near-term: a machine working through your checkout, your documentation or your pricing page is a user with no patience for interstitials, no ability to interpret an image of a price table, and no tolerance for information that exists only inside a script-rendered widget.

Accessibility work and clean server-rendered markup, long argued for on other grounds, now have a commercial one.

Access becoming a negotiation

The open assumption that anything crawlable is fair training material has broken down. Large publishers now sign licensing agreements, block specific AI crawlers, or both, and the pattern is spreading down to mid-sized sites.

For a business that is not selling access to its content, this is an opportunity rather than a threat.

As reference-grade publishers restrict access, the systems still need sources, and a site that is open, well structured and genuinely useful becomes relatively more valuable. Blocking crawlers is a decision with real trade-offs, and Managing AI Crawler Access works through them properly.

Measurement moving from clicks to mentions

Analytics built to count visits cannot see a brand named inside an answer the reader never clicked. That gap is now large enough that a category of tools exists to fill it, tracking how often and how favourably a brand appears across assistants.

The category is young and the tools disagree with each other, because they sample different prompts at different times and generated answers vary between runs. Use them as a trend instrument, never as a source of truth, and keep your own prompt log alongside whatever you buy.

Strong Signal

Answers that are not only text

Models that read images, video and audio are already deployed. What is still arriving is those capabilities becoming the default path through which a question gets answered: a spoken question producing a spoken answer, a photograph producing a product recommendation.

The preparation is unglamorous and mostly things you should already do. Real alt text rather than keyword padding. Transcripts for video, which serve both accessibility and extraction. Product information as text near the image rather than baked into it. Nothing here is a bet: it pays off under current conditions too.

Personalisation making a single ranking meaningless

Assistants increasingly carry memory of a user across sessions. Two people asking an identical question get different answers based on what the system knows about each of them.

If this deepens, the idea of a fixed position for a query stops being coherent, in the way that a single national ranking stopped being coherent once search went local. Measurement built on one prompt in one session becomes an anecdote.

The defensible response is to test in fresh sessions, test repeatedly, and treat variance as data about the system rather than noise to average away.

Structured access replacing scraping

There is visible movement toward sites offering machine-readable access deliberately rather than being scraped incidentally, from schema-rich pages through to emerging conventions for declaring what an assistant may use.

llms.txt belongs in this category, and it deserves a caveat: it is a proposed convention with real interest behind it and no confirmed universal support from the major systems. Adding one is cheap and harmless. Treating it as a ranking factor is not supported by anything published.

Speculation

These get discussed as though they were settled. They are not, and betting a strategy on them would be premature:

• Paid placement inside generated answers. Commercially plausible and technically straightforward. It would also directly undermine the trust that makes the answers useful, and no major platform has shipped it at scale. Worth watching, not worth planning around.
• A single dominant assistant. The current pattern is fragmentation, with distinct platforms serving distinct audiences and citing overlapping but different sources. Consolidation is possible. Nothing visible as of August 2026 points to it.
• The end of websites. Recurs every few years in some form. Assistants need sources, sources need to live somewhere, and the systems that cite most heavily are the ones sending traffic onward. The shape of a useful website is changing. Its necessity is not in evidence.

• Already happening observable now
  • Assistants that act on your site rather than only describe it
  • Access to content becoming a negotiation rather than a default
  • Measurement moving from clicks to mentions
• Strong signal directionally likely
  • Answers that are not only text
  • Personalisation making a single ranking meaningless
  • Structured access replacing scraping
• Speculation discussed as settled, and is not
  • Paid placement inside generated answers
  • A single dominant assistant
  • The end of websites

What To Do With This

The practical value of a lesson about the future is a short list of things that pay off under several plausible outcomes at once. Every item here is worth doing even if every prediction above is wrong:

1. Make your content machine-readable without javascript. Server-rendered HTML with real semantic structure serves agents, crawlers, screen readers and slow connections. No forecast required.
2. Keep your own measurement. A monthly log of the same prompts on the same platforms will outlive whichever tool you are paying for, and it is the only record that is genuinely yours.
3. Invest in what compounds. Original data, consistent entity presence and earned coverage matter under every scenario above. Tactics tied to one platform's current behaviour do not.
4. Decide your crawler policy deliberately. Whether you allow, restrict or license access, make it a decision with a stated reason rather than a default you never examined.
5. Re-read your own predictions. Write down what you expect to be true in twelve months, date it, and check it. Almost nobody in this field does, which is precisely why so much confident forecasting survives contact with reality.

Key Takeaways

• Sort every claim about the future into already happening, strong signal or speculation, and act only on the first two.
• Agentic systems, contested crawler access and mention-based measurement are visible now, not forecasts.
• Multimodal answers, deep personalisation and structured access conventions have momentum but are not settled.
• Paid placement, a single winning assistant and the death of websites are speculation and should be treated as such.
• The durable strategy is the set of actions that pay off under several futures: clean markup, your own measurement, and investment in what compounds.