GEO Trends in Fall 2026: A Citation Lives 11 Days, and the Agent Makes the Call

TL;DR

September 2026 brought three major studies, and they change work priorities. Schema and llms.txt barely affect whether an AI recommends you: after adjusting for site authority, the gain is 0.4 percentage points. A clear explanation of what you sell and to whom, plus consistent facts about the company across sources, make a difference of several times. A citation does not live long: half of all pages lose half of their mentions within 11 days of their peak. And your site has a new reader, an AI agent that checks price, availability and the option to place an order on its own.

Key facts

• Any schema markup adds 0.4 percentage points to the probability of a recommendation; FAQ schema and llms.txt have a negative effect.
• Brands with consistent facts across sources make the agent’s shortlist in 46.2% of cases, brands with conflicting facts in 2.8%.
• Half of all pages lose half of their citations within 11 days of their peak.
• 48% of the URLs that AI cites are mentioned only once a month.
• Expanding the prompt set from 10 to 100 cut the spread of the visibility estimate by 73%.
• Sites that block AI bots or depend entirely on JavaScript appear in answers in 3.1% of cases.
• A machine-readable way to place an order raises the chance of making the agent’s shortlist from 19.4% to 36.8%.

What changed in GEO in fall 2026?

GEO used to answer the question “how do we get cited?”. Now the question is different: will the company make the shortlist that the AI puts together, and can an agent verify its price, availability and terms? The focus is shifting from text to data about the company.

The most data came from a First Page Sage study published on October 2. It covers 4,213 commercial prompts in ChatGPT, Gemini and Claude, 657 tasks for AI agents and 1,089 brands across 14 industries. The study compared regular AI search with agent behavior separately, and the factor weights turned out to be different.

What drives a recommendation	Regular AI search	When an agent acts
Mentions in authoritative roundups	33.1%	21.2%
Fit with the customer’s need	9.6%	13.8%
Consistent facts across sources	5.2%	12.6%
Clear product explanation	3.4%	6.3%
Schema markup	0.9%	4.6%

This is each factor’s share of influence in the researchers’ model, not an exact law of the market. But the direction is clear: external mentions remain the main driver, and for agents the role of consistency and specifics grows sharply.

Do schema and llms.txt help you get into AI answers?

Barely. After adjusting for site authority, any schema markup added only 0.4 percentage points, and Organization schema 0.2. FAQ schema and llms.txt went negative, by 0.3 and 0.1 points. There is one exception: Product schema for online stores, plus 4.8 points.

The comparison among clear, well-explained sites is even more telling. A minimal structure without schema got 24.2% of recommendations, while the maximal setup with six or more markup types, FAQ and llms.txt got 23.9%. No difference.

Clarity, on the other hand, matters a lot. Brands that poorly explain what exactly they sell and who it is for received 11.2% of recommendations. Those that explain it well received 23.6%. That is twice as much, and it is the cheapest fix of all.

Why does AI recommend a competitor when your product is better?

Most often because the facts about you differ from source to source. In regular AI search, brands with very high data consistency received 34.9% of recommendations, those with very low consistency 5.9%. In agent tasks the gap is even wider: 46.2% versus 2.8%.

External consistency also matters more than order on your own site: 17.1% versus 13.7%. Directories, listings, profiles and publications should say the same thing as your website.

How long does a company mention last in AI answers?

A median of 11 days. Profound tracked 883,000 pages across seven systems over a year, which adds up to 1.19 million “page plus system” pairs. Half of the pages lost half of their citations within 11 days of their peak, and 78% went the same way within two weeks.

The second important number: from the first mention to a twofold drop takes 42 days on average. Pages cited by at least five systems are 5.5 times more likely to stay visible after eight weeks. And evergreen content, such as comparisons and guides, lives 1.4-1.6 times longer than news.

The takeaway is simple. “Publish and forget” no longer works. You need a cycle: monitor, notice the decline, update the facts, bring the content back into external sources.

How should you measure visibility in ChatGPT and Gemini?

With a broad set of prompts, not your ten favorite ones. Otterly collected 252,407 answers to 520 prompts across seven systems. With ten prompts, the visibility estimate swung between 39% and 71%. With a hundred prompts, the range narrowed to 51-60%, meaning the spread dropped by 73%.

• Brand mentions are more stable than links: only 12-19% of brands appeared on just one day out of thirty.
• Sources are highly fluid: 48% of 46,009 URLs were cited only once, and the top 1% of URLs accounted for 44% of all citations.
• Systems behave differently: the next day Perplexity keeps about 75% of yesterday’s sources, ChatGPT and Google AI Mode about 26%.

So the share of prompts where you are mentioned should become your main metric. Keep the list of sources separately, as a diagnostic: it explains what that visibility rests on.

Why does AI recommend different brands to two different customers?

Because the answer depends on who is asking. Profound analyzed 71,147 answers from ChatGPT, Claude and Gemini while varying the user’s income, age, gender and profession. Answers for different personas overlapped only about a quarter by brands and a fifth by sources.

Income had the strongest effect. High-income personas received 7.5 brand mentions on average versus 5.8 for low-income personas. That is why the question “what is our visibility in ChatGPT?” is too blunt. It is more useful to know whether you are recommended to a specific customer group in a specific situation.

What does an AI agent need to take a customer all the way to an order?

Facts a machine can read: price, availability, specifications and a way to place an order. Across 657 agent tasks in the First Page Sage study, the difference looked like this.

Product or service data	Machine-readable	Text only	Missing
Price	31.4%	22.8%	9.7%
Availability	34.2%	21.5%	8.1%
Specifications	30.6%	23.9%	12.4%
Booking or checkout	36.8%	19.4%	6.2%

This is the rate of making the agent’s shortlist. The biggest gap is in booking and checkout. But first the fact simply has to exist: the difference between “missing” and “text only” is bigger than between “text only” and “machine-readable”.

There is also a technical threshold. Brands that block AI bots or depend entirely on JavaScript appeared in only 3.1% of relevant prompts. How to check access is covered in our step-by-step checklist for getting into ChatGPT answers. Online stores also need a product feed. In September, Microsoft Advertising explicitly advised giving AI systems complete and up-to-date product information so they can see the whole range.

What should a business do in the next month?

Five steps, in order of impact.

1. Clarity. Rewrite your key pages so the first lines make it clear what you sell, who it is for and how much it costs.
2. Consistency. Check your name, addresses, prices and list of services on your site, in directories, profiles and publications. Fix any mismatches.
3. Access. Check whether AI bots can see your site and whether it reads without JavaScript.
4. Facts for action. Put price, availability, terms and the way to order on the page, and set up a product feed if you run a store.
5. Measurement. Build a set of one hundred customer prompts and check once a month how many of them mention you.

How these steps fit into one system is shown in the Haivoronskyi framework.