How to show up in ChatGPT: a developer's playbook

TL;DR

• On 8 and 14 August 2026, in two discrete steps rather than a slow drift, ChatGPT cut Reddit citations by 73.4% while its total citation volume rose 3.5%, according to Otterly, published 27/08/2026. What replaced Reddit was official brand pages, reference sources and public government databases.
• Citation rates diverge sharply by sector. In Similarweb's series from June 2025 to May 2026, travel and hospitality sit near 23%, automotive near 20%, and professional services below 4%.
• For a developer, the work left is not producing more content. It is publishing as text the project facts that currently exist only inside an image.

The week the GEO playbook stopped working

For nearly two years the advice sold to anyone who wanted to appear in AI answers was roughly identical: show up where people talk. Forums, communities, Reddit threads, expert replies. The logic was sound, because the engines really did pull discussion at a high rate.

In August 2026 that changed inside a single week. Promptwatch recorded reddit.com's share of ChatGPT Search citations collapsing from about 4% to about 0.5% on 14 August, and Axios reported the move on 20/08/2026. Otterly, tracking 16 brand reports across unrelated industries, measured a 73.4% decline and published on 27/08/2026. The number that matters sits in the other half of the same measurement: ChatGPT's total citation volume rose 3.5% over the period. The engine did not cite less. It cited elsewhere.

Otterly breaks down what rose in Reddit's place. Official brand pages and competitor sites gained in 10 of the 15 reports with comparable data. Reference and encyclopedia sources gained in 4 of 5. Government and regulatory databases entered the mix. Otterly's own reading is blunt: the engine traded discussion and reporting for canonical and first-party sources.

Worth recording that the measurement firms disagree on the size of the drop, and the disagreement is itself information. Promptwatch says 4% to 0.5%. Otterly says 73.4% across a monitored brand set. Nobody holds the official figure, because there is no official figure. Anyone who wants to know where they stand has to measure their own brand.

Why this lands hard on real estate

The sector entered this shift badly positioned. In May we published the AI search blackout that put luxury real estate last among every industry measured for AI visibility. The cause identified there was structural: an industry that built discovery on print, galas and broker relationships leaves almost no textual trail for an engine to pull.

Similarweb's data confirms the asymmetry. Between June 2025 and May 2026 the US citation rate climbed from roughly 1.6% to roughly 6.8%, but the average hides the spread between categories: travel and hospitality near 23%, automotive near 20%, professional services below 4%. The sectors that surface are the ones that already published specification in text, with price, address, hours and availability. The ones that vanish publish adjectives. We made the adjacent argument about operations in AI in real estate: hype versus the invisible engine.

Kevin Indig's research published on Growth Memo on 25/03/2026, drawn from 98,000 citations across 1.2 million ChatGPT answers in seven verticals, sets the effort benchmark. Thirty domains hold 67% of citations within each vertical. Eighty-five per cent of the pages an engine retrieves are never cited. And a page ranking first on Google carries a 43.2% chance of being cited. That last figure dismantles the claim that search optimisation died. It became the entry ticket rather than the destination.

The thesis: a developer's site has a fact problem, not a content problem

Open the website of almost any high-end developer and count the verifiable facts that exist as a sentence. The address usually sits inside a render caption. The completion date lives in a PDF brochure. The unit count appears in a carousel that rotates on its own. The architect's name is a signature inside an image. All of it is invisible to the layer that decides who gets cited.

Here is the uncomfortable part. Developers hold a base of public, checkable facts that most sectors would envy, and barely use it. Planning approvals, permits, land registry entries, condominium filings, unit mix by typology, floor areas, delivery schedule, city, district. That is registry-grade documentation, verifiable by a third party, precisely the class of source ChatGPT started preferring in August. The industry buries it under campaign language and then hires a GEO consultancy to add structured markup to a page that asserts nothing.

Which is why much of what is sold today as AI visibility strategy fails to deliver: it applies schema to a page with nothing to mark up. The correct order runs the other way. First the fact in text, then the structure that describes it.

The playbook, in six steps

1. Measure before you hire anyone. Write ten questions a buyer would genuinely ask and run the same ten through ChatGPT, Gemini, Perplexity and Google's AI Mode on the same day. Best developers in a named city. Who built a named project. Trusted developer in a named district. Record who appears and, more importantly, which domain the citation came from. Without that baseline no agency proposal is assessable, because there is nothing to compare it against.

2. Turn images into sentences. Every project page needs, as text the browser renders: full address with district and city, unit count, typologies and floor areas, expected completion, the architecture and landscape practices by name, and the permit or filing reference. Not for aesthetics, but because that is what a machine can lift and attribute.

3. Make the company page answer who the company is. Year founded, number of projects delivered, cumulative built area, cities of operation, company registration number, industry memberships. That is the canonical first-party page the August shift started rewarding, and on most sites it is the vaguest page of all.

4. Feed the databases the machine treats as a registry. Business profiles on Google and Bing Places, since ChatGPT's local answers route through them. Trade association pages. Portals where the company exists as a listed entity with a profile rather than an advert. Consistency of name, address and registration number across all of them counts for more than volume, because a mismatched record is what makes an engine distrust the entity.

5. Write one deep piece per subject instead of twelve shallow ones. In Indig's measurement, pages above 20,000 characters average 10.18 citations, against 2.39 for pages under 500. One complete page about the district a company builds in beats twelve thousand-character posts about the same district.

6. Repeat the measurement monthly. August 2026 proved the citation base can move in a single day, and it moved twice in one week. An annual audit measures a world that has already ended.

Three expensive mistakes right now

Buying forum mention packages. That is precisely the market that evaporated in August, and it was already under suspicion, with reporting through early 2026 on firms gaming Reddit to influence AI answers. Buying it now means paying rent on an asset whose lease just expired.

Blocking the crawlers without deciding to. Similarweb found that news publishers blocking language-model access rank far below what their readership would justify. Blocking is a legitimate choice, but it is a choice to disappear, and it should be made under that name.

Confusing citation with sale. In March 2026 a large Brazilian portal shipped property search inside ChatGPT, reported by Exame on 17/03/2026, and Zillow sits among the brands Otterly tracks. In residential, assistants are being wired into aggregator inventory. Being cited in an answer that ends on a portal listing generates a lead for the portal. Appearing is necessary and not sufficient, and the measurement has to keep the two apart.

Frequently asked questions

Do I need to hire a GEO agency to show up in ChatGPT?

Measure first. Run the ten questions from step one and look at which domains the citations in your sector and city actually come from. If they come from portals and company pages, the work is owned content and registry hygiene, and much of it is internal. Hiring without a baseline means signing a contract whose result nobody can verify.

How long before a developer starts getting cited?

No credible published timeline exists with a stated methodology, and anyone promising one deserves suspicion. What Indig's measurement indicates is the path: since a page ranking first on Google has a 43.2% chance of being cited, citation gains tend to follow organic ranking gains, which in real estate are counted in months.

Is publishing an llms.txt file worth it?

It is cheap and harmless, so it can go in. But no major engine has publicly confirmed using the file as a citation factor, and it replaces none of the six steps above. Treating llms.txt as a strategy is dressing the facade of a building with no address.