ChatGPT Rebuilt Its Search Tool, I Read The New Language It Speaks

Aug 26, 2026 02:00 AM - 2 hours ago 1

On August 16, erstwhile ChatGPT ran a web hunt for me, the instrumentality telephone was JSON.

{"system1_search_query":[{"q":"site:intercom.com Fin AI Agent pricing 2026"}]}

On August 20, I asked the aforesaid mobility connected the aforesaid account. The instrumentality telephone now looks for illustration this.

Somewhere successful those 4 days, OpenAI replaced the JSON pinch a compact query connection of its own. And search_queries, the metadata section that mirrored the fan-out, the section my past article told you to publication and the section FanoutFox parses, is gone from the payload entirely.

This is Part 4 of the bid that started pinch how ChatGPT picks sources and astir precocious recovered the shortlist it writes earlier it searches. This clip I ran the caller format crossed 8 questions, spanning commercial, beingness goods, local, news and finance, and mapped what the connection tin say.

It says a batch much than the aged one.

Five ideas, each 1 built connected the last.

Every statement is simply a search, and the fields are readable.
The 3rd section is simply a freshness window, and it changes pinch what you ask.
Search is simply a group of verticals, and your contented type decides which 1 fires.
The domain slot intends ChatGPT visits sites it already knows.
Some answers are becoming widgets, and there’s nary citation to triumph successful them.

Everything present is publication disconnected my ain browser postulation pinch a azygous ChatGPT Plus relationship from a azygous location.

Idea 1: Every Line Is A Search

A instrumentality telephone is now a artifact of lines, 1 hunt per line, fields separated by pipes. The shape that repeats is simply a telephone type, the query, a number, and sometimes a domain.

fast|Zendesk AI agents pricing 2026|30|zendesk.com

fast is the telephone type. The query sounds for illustration the aged fan-out queries, marque names the exemplary wrote successful itself, a year, an intent connection for illustration pricing. The number and the domain are new, and they’re the absorbing part, because they’re parameters the aged format had obscurity to put.

Each artifact closes pinch a directive line, length|long, length|medium aliases length|short. This 1 is the subsister of the aged format. The JSON astatine the apical of this article had "response_length":"long" successful the aforesaid position, truthful it’s been renamed alternatively than added. It sounds arsenic an instruction for really overmuch matter to bring backmost per result, and the values way the job. agelong fired connected the heavy merchandise comparison, mean connected astir lookups, and short connected a last telephone that was verifying 2 venue names. If that reference is right, astir searches propulsion a bounded excerpt of your page alternatively than the page, 1 much logic the condemnation that answers the mobility needs to beryllium wherever an excerpt starts.

The fields aft the query are optional. Some lines are only a query, immoderate adhd the number, and the domain appears erstwhile a tract is being targeted.

Practical AI SEO/GEO Tip

Ask the questions your buyers inquire and cheque the brands written by ChatGPT. The marque names successful those queries are the competitory group ChatGPT already holds for your category, and the domains successful the slots are the sites it intends to visit.

Make judge your tract will past the probing (I person written astir this earlier and made a fewer tools, truthful cheque them out).

Idea 2: The 3rd Field Is A Freshness Window

Here’s the aforesaid section crossed 5 lines from the questions I asked connected August 20.

QuestionLineNumber

Is Nvidia overvalued	fast\|NVIDIA NVDA banal value August 20 2026...\|2	2
Premier League weekend	fast\|Premier League results...\|7\|premierleague.com	7
Best AI unrecorded chat software	fast\|Zendesk AI agents pricing 2026\|30\|zendesk.com	30
Nvidia net guidance	fast\|NVIDIA latest net guidance...\|90\|investor.nvidia.com	90
Live chat, the Reddit line	fast\|best AI unrecorded chat support...reddit\|365\|reddit.com	365

A banal value gets two. Football results get seven. Commercial merchandise investigation gets 30. Earnings guidance gets 90.

So, what connected world is this number?

I deliberation this number is simply a recency model successful days, matched to really accelerated the reply goes stale.

I haven’t seen immoderate spec for this, truthful this is my assumption.

Practical AI SEO/GEO Tip

On the commercialized questions I ran, each marque probe had a 30-day window.

If the days presumption is right, a pricing aliases comparison page that hasn’t changed successful a period is competing from extracurricular the default model connected precisely the queries wherever brands get compared.

Update the pages that reply buying questions wrong that cycle, pinch existent changes and visible dates.

Same maneuver goes to your cardinal pages and content. Check your prompts/keywords.

Idea 3: Search Is A Set Of Verticals

The telephone types I’ve captured truthful far, pinch the existent lines.

fast is the web search. It’s the successor to the full aged fan-out.

product is simply a catalogue lookup. It appeared connected robot vacuums and mattresses, and connected nary of the package questions I ran.

product|Roborock Saros 10R robot vacuum;Dreame X50 Ultra robot vacuum;Eufy E28 robot vacuum

Product names, semicolon-separated, pinch thing other connected the line. It lines up pinch the merchandise cards and merchant offers I recovered successful the commerce furniture backmost successful June.

For beingness goods, ChatGPT looks products up successful a catalogue arsenic good arsenic searching pages astir them, and beingness successful that catalogue is its ain game, abstracted from thing your blog does.

Practical AI SEO/GEO Tip

If you waste beingness products, cheque whether yours travel backmost pinch cards, offers and merchant names erstwhile ChatGPT recommends them. Cards mean the catalogue knows you. A matter mention pinch nary paper intends the contented crippled is the only 1 you’re in, and nary magnitude of page activity changes what the merchandise telephone returns.

business is simply a places lookup, and it takes a location. On a java question, it fired 3 times. Here are the first two.

business|Dubai, UAE|specialty java Dubai Marina;best java Dubai Marina business|Dubai, UAE||Blends Specialty Coffee;Roast Speciality Coffee, Marina;OLAB Coffee Shop JBR

The first telephone sends hunt phrases. The 2nd sends existent business names it sewage backmost from the first, and past verifies them 1 by 1 pinch quoted accelerated searches.

That’s the section type of the shortlist mechanics from the shortlist article, moving connected venues alternatively of brands.

If you tally a section business, the portion of optimisation is your entity successful the places data.

The lookup returns the name, the reviews, and the details, and the web searches only verify them.

Ask ChatGPT for your ain class successful your ain area and watch the 2nd business call.

If your venue’s sanction appears successful it, the lookup knows you and the follow-up searches are checking your details, truthful make judge what’s connected your tract matches what the listings say. If your sanction ne'er enters that call, your problem is the places data, and your website can’t hole it.

image is an image search. image|Dubai Marina specialty java shop|30. It fired alongside the places lookup.

genui_run calls a widget, pinch JSON arguments.

genui_run|epl_schedule|{"fn":"schedule","league":"EPL","date_from":"2026-08-21","date_to":"2026-08-24","locale_override":"en-GB"} genui_run|stock_chart|{"ticker":"NVDA","asset_type":"equity","market":"USA","locale_override":"en-US"}

A shot schedule and a banal chart, requested arsenic functions pinch typed arguments, not arsenic searches astatine all. More connected what that intends successful Idea 5.

One much item from the news capture. The domain section tin clasp a list. theguardian.com;skysports.com appeared connected 1 line, truthful a hunt tin beryllium scoped to respective publishers astatine once.

Idea 4: The Domain Slot

Until this week, erstwhile ChatGPT wanted to hunt 1 circumstantial website, it wrote the classical site:intercom.com successful beforehand of the query, the aforesaid usability hunt we SEOs often use…

The caller format has a slot for it instead. The past section connected a statement is simply a website, and that hunt only looks there. Checking 1 company’s tract is built into the instrumentality now, sitting adjacent to the query connected each line, alternatively than spelled retired successful the hunt words. Both styles still show up; immoderate lines support site: successful the words while others usage the slot.

What this intends successful believe hasn’t moved from the shortlist article. The exemplary still names brands earlier it fetches anything, past goes to their websites to cheque what they say.

What’s caller is really mean that measurement is. Visiting the website of a institution it conscionable named is simply a modular action connected each hunt line, and the exemplary fills the slot itself, from immoderate domain it already links to the brand.

If the domain successful its representation is incorrect aliases stale, the probe searches the incorrect place. I’ve watched that hap once: a probe aimed astatine a domain the marque doesn’t run, and the wide hunt rescued them.

On the query champion llm visibility instrumentality and what does each 1 cost, the exemplary wrote this probe:

site:profound.ai pricing AI visibility level Profound pricing 2026

Profound doesn’t tally connected profound.ai; they tally tryprofound.com. The probe returned zero pages from the pool. The marque survived anyway. The wide find hunt recovered the correct domain; tryprofound.com sewage cited astatine its pricing page, and Profound was named six times successful the reply body. So the incorrect slot costs them 1 of the model’s 4 searches.

Practical AI SEO/GEO Tip

Read the slot connected your ain captures and cheque the domain is 1 you really run. Rebrands, migrations, and country domains are the cases that nutrient a incorrect slot, because the exemplary fills it from memory, and representation lags. A wasted probe costs that marque thing connected the time I watched it, and I wouldn’t build a strategy connected the rescue repeating.

Idea 5: Some Answers Are Becoming Widgets

The genui_run lines don’t return pages. They return interface.

In the finance answer, the banal floor plan showed up arsenic a reference named stock_chart pointing astatine a widget hosted connected cdn.platform.openai.com, type 2.0, hydrated from a authorities payload. In the java answer, the exemplary emitted a genui directive successful the reply matter itself, a representation widget pinch pins wired to the businesses the places lookup returned.

So, for sports, stocks, places, and presumably much I haven’t caught yet, portion of the reply is simply a constituent OpenAI renders itself. A banal floor plan has nary nexus wrong it to win, and the shot schedule cites nary fixtures page. Whole classes of query are moving to surfaces wherever the citation game, the 1 each GEO instrumentality measures, doesn’t exist.

Practical AI SEO/GEO Tip

Run the queries you rank for and statement which travel backmost pinch a widget wherever the reply utilized to be. A genui_run statement successful the seizure is the tell. I can’t springiness you a maneuver for winning wrong a widget, because connected what I’ve seen, there’s thing successful 1 to win. What you tin do is cognize which of your queries person crossed over, and extremity counting citation opportunities connected reply types that nary longer connection any. If your postulation depends connected being the page that answers “what clip is the match,” the widget is your competitor now.

It’s the aforesaid telephone arsenic not optimizing for a keyword whose SERP is an AI Overview and a wall of different features.

The Reddit Situation Is Very Very Interesting

I want to locomotion you done this 1 measurement by step, keeping what I tin spot abstracted from what I deliberation it means.

Start pinch the statement itself, from the unrecorded chat capture.

fast|best AI unrecorded chat support Intercom Gorgias Zendesk Ada Tidio Crisp reddit|365|reddit.com

Read the query. Reddit is named successful the words, the domain slot is pointed astatine reddit.com, and the model is 365, wherever each marque statement successful the aforesaid telephone sewage 30. And the six marque names are successful location pinch it.

So, what’s going connected here?

I genuinely had a Joey moment. (Friends fans will get it.)

So this hunt didn’t inquire Reddit who the bully unrecorded chat devices are. It asked Reddit what group opportunity astir the six devices the exemplary had already picked.

The campaigner group came from memory, and Reddit was fetched for sentiment astir it.

What came backmost is the adjacent tangible fact.

Eighty-four of the 221 entries successful that conversation’s retrieval excavation were Reddit threads, much than from immoderate different source. The speech had nary open-and-read step, truthful what entered the model’s discourse was 84 thread snippets, the title and excerpt the hunt returns, alternatively than 84 threads publication successful full.

Then the reply went retired pinch each citation bound to vendor pages. Zero of the 84 sewage credited, successful the 1 speech wherever I’ve checked citations against the excavation statement by line.

The java seizure past did it again, harder.

fast|site:reddit.com Dubai Marina specialty java Dubai|3650|reddit.com

Both mechanisms astatine once, site: successful the words and the domain successful the slot, pinch a 3650-day window. Ten years of threads, connected a mobility wherever news sites sewage 7 days.

Why Reddit would get a decade connected a section query, I honestly don’t know, but I person a fewer thoughts connected this below. One hint astir why a wide model fits the question: I asked my AI supplier to hunt Reddit for Dubai Marina java threads, and it recovered immoderate arsenic aged arsenic 12 years. That’s the archive a 3,650-day model reaches.

The ligament itself can’t corroborate ages either way. Almost nary of the fetched Reddit entries successful my seizure carried a date, truthful I can’t show you really aged the threads ChatGPT really pulled were.

Promptwatch’s floor plan from August 18 has Reddit’s stock of ChatGPT citations collapsing from astir 3.83% to nether 1% successful mid August. Their conjecture was a root action change. A citation sheet only sees the vanished answer. The ligament adds the steps earlier it: the deliberate request, the bulk delivery, and the zero credits.

Based connected what I’m seeing, this is what I deliberation is happening. I can’t corroborate it yet, truthful dainty this arsenic my mentation of the evidence, not a definitive conclusion.

Content sitting successful the model’s discourse tin style what it writes. (There is often a bias.)
This exemplary fetched sentiment astir its shortlist instantly earlier penning opinions astir its shortlist, past attributed those opinions to nobody.
I deliberation Reddit has go an invisible input, feeding the verdicts while the credits spell to vendor pages.
That would fresh what I recovered successful Part 1, wherever facts came from charismatic pages, and verdicts came from 3rd parties, pinch Reddit main among them. If that section of labour survived and the crediting stopped, the power is still there, and the receipts are gone.
The different imaginable reference is that the snippets are fetched and past ignored, and the year-wide model is simply a leftover habit. Nothing successful the postulation separates the two. The fetching and the zero credits are facts I tin show you. The power is my interpretation, and the ligament cannot beryllium power without citation.

Now let’s unpack this a small further and look astatine the wider discourse astir why OpenAI mightiness person made this change.

Anti Manipulation And Spam Deterrence

This could beryllium a measurement to trim the effect of group gaming Reddit for AI SEO/GEO. If ChatGPT is deliberately looking crossed 1 to 10 years of Reddit history, caller manipulation becomes a overmuch smaller signal.

There are efficaciously 2 filters. The campaigner excavation appears to beryllium selected first, past Reddit is utilized to stitchery opinions aliases validate those candidates. That makes simply flooding Reddit pinch marque mentions acold little useful.

There whitethorn besides beryllium a deterrence effect. Citation studies show Reddit visibility falling sharply, which makes marketers deliberation Reddit has mislaid its value. Less perceived worth intends little spam.

Compliance Or Licensing

For context, OpenAI and Reddit signed a information licensing woody successful 2024, truthful the information entree itself is paid for and public. What that woody says astir attribution, I can’t spot from the wire. Another possibility, and it’s speculation, is that OpenAI still wants the worth of Reddit information without leaning connected Reddit arsenic an obvious, straight cited source.

Instead of surfacing Reddit pages prominently successful answers, Reddit could beryllium utilized further upstream arsenic an input for campaigner evaluation, sentiment gathering, aliases different soul retrieval processes.

Practical AI SEO/GEO Tip

If Reddit has been portion of your AI visibility work, count the chips yourself. Run your queries, and cheque whether the threads you invested successful still get credited connected them.

Does this alteration make your Reddit campaigns useless? Absolutely not.

If you’ve been doing this for a while, genuinely engaging pinch Reddit communities and building a brand, particularly if you survived Reddit’s caller spam purge, transportation on. Two channels past this alteration either way. Years of threads provender the training information that builds the representation shortlist successful the first place, and the humans reference those threads ne'er stopped being buyers. What’s collapsed is the measurable citation payoff, the chips.

If your strategy was simply spamming Reddit to manipulate AI visibility, though, this is bad news for you.

Remember 1 thing. A thread that gets fetched but ne'er cited isn’t a measurable win. Whether it still sways the reply is the unfastened mobility above.

What Breaks, And What I’d Do

The aged look is dead. My past article said to unfastened DevTools and hunt the consequence for queries. That section nary longer exists. The queries are still there; they’ve moved into the web.run tool-call messages arsenic the pipe-delimited lines above. FanoutFox sounds the aged section too, and I’ll beryllium patching it.

The caller two-minute check. Open ChatGPT successful Chrome pinch DevTools connected the Network tab. Ask the commercialized mobility your buyers ask. Filter for conversation, click the biggest response, the speech payload, not the mini stream_status and textdocs rows that stock the name. Open its Response tab, property Cmd+F wrong that pane, and hunt for fast|. The panel-wide hunt drawer will drown you successful JS bundle hits, truthful hunt wrong the 1 consequence instead. Read the lines. You’re looking for 3 things. Whether your marque is successful the queries, whether your domain is successful the slot, and what model your class runs on.

That cheque is manual, and it gets aged connected the 3rd answer. FanoutFox, the free Chrome hold I built for this series, sounds the aforesaid convention information and does the fetched-versus-cited counting for you, per domain, per answer. The converts file is the Reddit cheque from this article arsenic a sorted table. Its fan-out paper sounds the aged search_queries field, truthful that paper is quiet connected caller conversations until the spot for the tube format ships, and the remainder of the panel, sources, citations, and the domain radar, is unaffected by the change.

This is the 3rd format alteration I’ve documented since June, and I could put a four-day model connected this 1 only because I support dated snapshots of what the postulation exposes and diff them aft each seizure session. If you do this benignant of work, support dated copies of everything. What was existent connected a Sunday was gone by Thursday.

More Resources: