What works in AI search in late 2026:
my verdict on OtterlyAI's GEO experiments

Which GEO tactics actually work in AI search right now?

Some do and some backfire. In OtterlyAI's October 2026 experiments, LinkedIn Pulse articles, real content updates and honest ranked lists earned AI citations. Bought Reddit activity did too, until the community was banned and the citations fell to zero. My advice is to copy the patient tactics and skip the shortcuts. Test everything on enough prompts to trust the result.

GEO stands for generative engine optimisation. It is the work of getting a brand named or cited in AI answers. Plenty of people sell GEO advice. Few publish tests. OtterlyAI ran its tests on live sites and shared the numbers. This essay takes each test in turn. It sets out what OtterlyAI found and what I would do with it.

Where do these numbers come from?

OtterlyAI sells AI search monitoring software. In October 2026 it presented a set of GEO experiments ("GEO Experiments 2026 Q4 Updates"). It ran them on its own sites and on tracking data from seven AI engines. That makes the research useful and also worth reading with care. Some tests are small. Some were run by a company with a product to sell. I give my verdict on each one below.

The table sums up each tactic. The sections after it give the detail.

Reddit: do not fake it

OtterlyAI posted daily into four subreddits for 60 days. Some posts got bot comments and bought upvotes. Citations climbed to around 280 a day in mid-June. Then the community was banned on 22 June. After the ChatGPT update on 8 August, citations flatlined.

OtterlyAI's own conclusion is that content and replies get cited. Upvotes and subscriber counts do not matter. The bigger lesson is in its chart of Reddit's reach. Reddit's share of ChatGPT's social citations in the US fell from close to 4% to under 1% on 7 August. A tactic that one platform update can wipe out is a gamble.

My verdict: take part in real communities with a real account. Never buy upvotes or comments. It breaks Reddit's rules and the gains do not last.

LinkedIn Pulse: worth it if you already use LinkedIn

Over 30 days, OtterlyAI published 240 LinkedIn posts and Pulse articles across four pages. Pulse articles from a personal profile took 72.9% of the citations. Company Pulse articles took 25.2%. Company feed posts took 1.9%. Personal feed posts got none.

None of these citations came from ChatGPT, Claude, Gemini or Google AI Overviews.

There is a catch. None of the citations came from ChatGPT, Claude, Gemini or Google AI Overviews. They came from Copilot, Perplexity and Google AI Mode. The test took 34 hours of work in a month.

My verdict: if you already write on LinkedIn, turn your best posts into Pulse articles under your own name. Do not start LinkedIn just for AI search.

Freshness: update for real, then date it

On 4 August 2026, OtterlyAI changed only the "Last updated" date on 10 articles. The content stayed the same. Over the next 30 days their citations rose 31.5%, from 4,800 to 6,312. Two control articles fell 10.6%.

Small sample. 10 updated articles and 2 controls.

Read that with care. It is 10 articles and 2 controls. The biggest jump was ChatGPT, from 37 citations to 293, which starts from a small base. The engines also moved in different directions. Google AI Mode fell 20% and Gemini fell 37%.

OtterlyAI itself says changing the date alone was a test. It does not recommend it as a tactic.

My verdict: keep important pages current. Fix what is out of date. Then let the visible date show that work. A new date on old content misleads readers.

AI-written content: use AI with an editor

OtterlyAI ran three sites on the same topic. One was written by people, one by AI and one was half and half. After four months the mixed site led on clicks, impressions and AI citations. The human-written articles were short and shallow, which OtterlyAI admits. In an earlier test, 1,000 one-click AI articles sank two sites within weeks.

My verdict: AI drafts plus a human editor who adds facts, sources and judgement is a fair way to work. One-click AI content is still a risk.

Listicles: publish your own ranking, honestly

In OtterlyAI's September 2026 study of over 519,000 classified citations, 66.6% went to ranked lists of the "best 10 X" kind. Head-to-head comparisons took 7.9%. How-to guides took 7.1%. Home pages took 3.4% and product or service pages took 2.7%. Brands publishing about their own category were the biggest single group of cited publishers.

The five formats shown. Other formats make up the rest.

Each engine differs. Claude, Perplexity and Gemini cite lists around three quarters of the time. Google's AI features lean more on Reddit and YouTube.

My verdict: a useful, honest ranking of your category can earn citations. A list that puts you first with no reason will not convince readers.

What did not work

OtterlyAI lists four tactics that showed no effect in its tests. The first is an llms.txt file. The second is a glossary hub. The third is schema markup for ChatGPT and Gemini. The fourth is image alt text and file names.

I would still keep alt text, because people who use screen readers need it. I would not expect any of these to move AI citations alone. My essay on what an AI visibility report checks explains why a missing llms.txt file shows as optional there.

Small prompt sets give misleading scores

OtterlyAI tracked 520 prompts across seven engines and collected 252,407 answers. With only 10 prompts, a brand's coverage swung between 39% and 72%. With 50 prompts the range was 48% to 63%. With 100 prompts the range narrowed to between 50% and 61%.

My essay on how prompt wording changes AI visibility scores points the same way. A score says as much about the prompts as about the brand. For what can and cannot be counted at all, see measuring AI visibility.

Free Q4 checklist

1. Track at least 100 prompts, split by engine.
2. Report brand coverage as the main number. Use citations to see which pages are doing the work.
3. List your 10 most important pages. Update what is out of date, then change the visible date.
4. Publish one honest ranking of your category, with clear reasons for each place.
5. If you already use LinkedIn, turn your two best posts into Pulse articles under your own name.
6. Answer real questions in Reddit communities from your own account. Buy nothing.
7. Check which engines your buyers use. Plan for those engines instead of the average.

Getting help

To track prompts and citations across engines, an AI visibility tool such as LLM Scout does the counting. For hands-on work on content and technical fixes, Orchid Box is an AI search agency. If you want an outside view first, the AI visibility audit on the hire page starts with these checks.

Frequently asked questions

Does changing the date on a page get more AI citations?

In OtterlyAI's test of 10 articles, citations rose 31.5% after only the date changed. The sample is small and OtterlyAI calls it a test. Update the content first, then change the date.

Should I buy Reddit upvotes to appear in ChatGPT?

No. OtterlyAI's subreddit was banned and its citations fell to zero. Upvotes did not drive citations. Content and replies did.

Does an llms.txt file help AI search?

It showed no effect in OtterlyAI's tests. Google says you do not need AI text files to appear in AI Overviews or AI Mode. It does no harm. Do not expect results from it.

How many prompts should I track?

At least 100. With 10 prompts, brand coverage swung between 39% and 72% in OtterlyAI's data.

Do LinkedIn posts get cited by ChatGPT?

Not in OtterlyAI's test. Its LinkedIn citations came from Copilot, Perplexity and Google AI Mode. Pulse articles from a personal profile took 72.9% of them.

Sources

1. OtterlyAI, GEO Experiments 2026 Q4 Updates: What We Tested, What Failed, What Actually Works (presentation, October 2026)
2. OtterlyAI, GEO experiments hub: otterly.ai/blog/geo-experiments
3. OtterlyAI, AI Search Index industry reports: insights.otterly.ai/industry-reports
4. Google Search Central, AI features and your website: developers.google.com/search/docs/appearance/ai-features