What actually works in GEO: two hacks died this summer, but one lever survives every update

I hear one question in almost every conversation, in some form or other: "What's the most important GEO lever right now?"

I get why people ask. People want the one trick that works today. Build lists. Show up on Reddit. Add tables. Something they can kick off on Monday.

The problem is that the answer to that question has a shelf life of a few months, sometimes a few days. August 2026 proved that more clearly than almost anything before it. So here is a look at both sides: what just broke, and what doesn't.

At a glance

• What happened: Two of the year's most popular GEO tips, listicles and a Reddit presence, lost most of their effect within a few days in August 2026.
• Listicles: After the ChatGPT update on August 6, their share of citations fell from 15.8 to 7.8 percent, according to Peec AI. Vendor product pages overtook them.
• Reddit: In mid-August, Reddit's share of ChatGPT citations dropped from 3.83 to 0.52 percent, according to Promptwatch. That is a decline of 86 percent. Nobody knows the cause.
• What survived: Content that is current, complete, and backed by evidence, and pages that actually contain the answer.
• The lever: Content that answers real customer questions. Formats and platforms are a thermometer, not a strategy.

Why don't listicles work in ChatGPT anymore?

Because the new ChatGPT model would rather get facts straight from the vendor than from third-party rankings. Early this year, the advice was simple: if you want ChatGPT to cite you, build listicles. "The 10 best tools for...," "7 vendors compared." For a long time, the citation data backed that up.

Then came August 6, 2026. That day, OpenAI moved Free and Go users to its new GPT-5.6 model and retuned the Plus and Pro models. Peec AI measured what happened to citations the same day:

• Listicles fell from 15.8 to 7.8 percent of all citations. More than half gone, in a single day.
• Comparison pages lost about a third, dropping from 9.1 to 6.2 percent.
• Vendor product pages reached 16.4 percent, which put them ahead of the lists.

If you rebuilt your content plan around lists this spring, you now have six months of work sunk into a format that one model update devalued overnight.

The lesson: a format is not a lever. It is a snapshot of what a model happens to prefer right now.

What happened to Reddit in ChatGPT?

Reddit practically vanished as a ChatGPT source in mid-August, and nobody knows exactly why. Building a forum presence, seeding threads, joining discussions: that was the second big GEO tip of the spring.

According to Promptwatch data reported by Search Engine Land on August 19, Reddit accounted for 3.83 percent of ChatGPT citations between July 18 and August 7. Between August 14 and 17, that share was down to 0.52 percent. An 86 percent drop within days, with no official announcement.

The uncomfortable part is that there is no explanation. Promptwatch itself says the data shows when it happened, not why, and explicitly calls the results provisional. Maybe a change in how sources are selected, maybe a side effect of the new model.

This is what a dying hack feels like from the inside: at some point you see fewer mentions, and you have no idea what is going on.

The lesson: if your visibility depends on one platform, it depends on something OpenAI can change on any given Thursday.

Which content survived the updates?

Content with substance survived: content that is current, complete, and backed by evidence. Not every list crashed, and not every source lost ground.

Back in February 2026, Seer Interactive analyzed more than two million ChatGPT citations and found an early listicle decline of about 30 percent month over month. What stood out was which lists kept getting cited: the ones kept up to date, genuinely complete with 10 to 20 entries instead of a token five, with disclosed methodology and verifiable data.

In August, the winners were the pages that actually contain the answer: vendor pages with real specs, real prices, and real use cases.

The pattern behind it never changes. The models keep getting better at finding the source that truly answers a customer's question. Every update moves them a little closer to what a good advisor would do: skip reciting someone else's ranking and go straight to the vendor who should know.

The lesson: no model update devalues substance. It devalues shortcuts.

What actually works in GEO?

Content that answers real customer questions. I know that sounds unexciting. It is the most boring sentence in all of GEO. It is also the only one I have never had to take back since I started doing this work.

Here is what that means in practice, in three parts.

Start with the questions, not the format

What do your customers ask AI before they contact you? Not your keywords, but the real questions: "Is X worth it for a 50-person company?" "What does Y realistically cost?" "How is A different from B?" Each of those questions deserves a page that answers it completely. You can measure which questions matter for your brand, and that is exactly where a GEO audit starts.

Write the answer down. All of it

Numbers, prices, limits, prerequisites, examples. No "measurably more efficient," no "tailored solutions." The pages that won in August were the ones with the facts on them. That is not a trick. It is the absence of one. I describe what that kind of writing looks like on my SEO and GEO copywriting page.

Keep it current and verifiable

Add a date, disclose your methodology, name your sources, update regularly. That is exactly what separated the lists that survived from the ones that died. And it happens to be exactly what a skeptical human needs before they believe you. To keep it from resting on one person, build it into your team's routines. That is what content enablement is for.

To be fair: watching formats and platforms still makes sense. As a thermometer. Not as a strategy.

Conclusion

Two hacks, one summer, both gone. Anyone who chased them lost more than the time they invested. They also lost the time that never went into content built to last.

What remains is inconvenient and boring: content that genuinely helps your customers decide. Backed by evidence instead of claims, kept current instead of built once and forgotten.

Hacks expire. Answers don't.

Ready for a GEO check?

Wondering whether GEO matters for your business and whether it would pay off? In a free 15-minute GEO quick call, we'll look at what GEO could do for your brand, what role AI search plays in your industry, and where it makes sense to start. This post first appeared in the GEO-Briefing, my LinkedIn newsletter on AI visibility.

→ Subscribe to the GEO-Briefing on LinkedIn (in German)

Frequently asked questions

Do listicles still work for GEO?

Not the way they used to. After the ChatGPT update on August 6, 2026, the listicle share of citations fell from 15.8 to 7.8 percent, according to Peec AI. Lists that stay current, complete, and backed by disclosed methodology still get cited. Thin rankings with only a few entries lose.

Why does ChatGPT barely cite Reddit anymore?

Nobody knows yet. According to Promptwatch, Reddit's share of ChatGPT citations fell from 3.83 to 0.52 percent in mid-August 2026. The data shows when it happened but not why, and Promptwatch considers it provisional. The one certainty: visibility that depends on a single platform can disappear at any time.

What content does ChatGPT prefer to cite?

Pages that actually contain the answer to a specific question. After the August update, vendor product pages reached 16.4 percent of citations, putting them ahead of listicles. What counts are hard facts like specs, prices, and use cases, plus freshness and verifiable sources.

What is the most durable GEO lever?

Content that answers your own customers' real questions completely and with evidence. Formats and platforms change with every model update. Substance doesn't. Watching formats still makes sense, but as a measuring tool, not as a strategy.