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

On August 16, when ChatGPT ran an online seek for me, the device name was JSON.

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

On August 20, I requested the identical query on the identical account. The device name now appears to be like like this.

quick|Intercom Fin AI agent pricing 2026 stay chat help|30|intercom.com
quick|Gorgias AI Agent pricing 2026 buyer help|30|gorgias.com
quick|Zendesk AI brokers pricing 2026|30|zendesk.com
quick|greatest AI stay chat help Intercom Gorgias Zendesk Ada Tidio Crisp reddit|365|reddit.com
size|lengthy

Someplace in these 4 days, OpenAI changed the JSON with a compact question language of its personal. And search_queries, the metadata subject that mirrored the fan-out, the sector my final article advised you to learn and the sector FanoutFox parses, is gone from the payload fully.

That is Half 4 of the collection that began with how ChatGPT picks sources and most not too long ago discovered the shortlist it writes before it searches. This time I ran the brand new format throughout eight questions, spanning industrial, bodily items, native, information and finance, and mapped what the language can say.

It says much more than the previous one.

5 concepts, every one constructed on the final.

1. Each line is a search, and the fields are readable.
2. The third subject is a freshness window, and it modifications with what you ask.
3. Search is a set of verticals, and your content material kind decides which one fires.
4. The area slot means ChatGPT visits websites it already is aware of.
5. Some solutions have gotten widgets, and there’s no quotation to win in them.

Every thing right here is learn off my very own browser visitors with a single ChatGPT Plus account from a single location.

Thought 1: Each Line Is A Search

A device name is now a block of strains, one search per line, fields separated by pipes. The sample that repeats is a name kind, the question, a quantity, and generally a site.

quick|Zendesk AI brokers pricing 2026|30|zendesk.com

quick is the decision kind. The question reads just like the previous fan-out queries, model names the mannequin wrote in itself, a yr, an intent phrase like pricing. The quantity and the area are new, and so they’re the fascinating half, as a result of they’re parameters the previous format had nowhere to place.

Every block closes with a directive line, size|lengthy, size|medium or size|quick. This one is the survivor of the previous format. The JSON on the prime of this text had "response_length":"lengthy" in the identical place, so it’s been renamed somewhat than added. It reads as an instruction for the way a lot textual content to carry again per end result, and the values observe the job. lengthy fired on the deep product comparability, medium on most lookups, and quick on a remaining name that was verifying two venue names. If that studying is correct, most searches pull a bounded excerpt of your web page somewhat than the web page, another reason the sentence that solutions the query wants to take a seat the place an excerpt begins.

The fields after the question are elective. Some strains are solely a question, some add the quantity, and the area seems when a web site is being focused.

Sensible AI search engine optimisation/GEO Tip

Ask the questions your patrons ask and test the manufacturers written by ChatGPT. The model names in these queries are the aggressive set ChatGPT already holds in your class, and the domains within the slots are the websites it intends to go to.

Make sure that your web site will survive the probing (I’ve written about this earlier than and made a couple of instruments, so test them out).

Thought 2: The third Discipline Is A Freshness Window

Right here’s the identical subject throughout 5 strains from the questions I requested on August 20.

A inventory worth will get two. Soccer outcomes get seven. Industrial product analysis will get 30. Earnings steerage will get 90.

So, what on earth is that this quantity?

I believe this quantity is a recency window in days, matched to how briskly the reply goes stale.

I haven’t seen any spec for this, so that is my assumption.

Sensible AI search engine optimisation/GEO Tip

On the industrial questions I ran, each model probe had a 30-day window.

If the times assumption is correct, a pricing or comparability web page that hasn’t modified in a month is competing from exterior the default window on precisely the queries the place manufacturers get in contrast.

Replace the pages that reply shopping for questions inside that cycle, with actual modifications and visual dates.

Similar tactic goes to your key pages and content material. Test your prompts/key phrases.

Thought 3: Search Is A Set Of Verticals

The decision varieties I’ve captured up to now, with the true strains.

quick is the online search. It’s the successor to the entire previous fan-out.

product is a list lookup. It appeared on robotic vacuums and mattresses, and on not one of the software program questions I ran.

product|Roborock Saros 10R robotic vacuum;Dreame X50 Extremely robotic vacuum;Eufy E28 robotic vacuum

Product names, semicolon-separated, with nothing else on the road. It strains up with the product playing cards and service provider presents I discovered within the commerce layer again in June.

For bodily items, ChatGPT appears to be like merchandise up in a list in addition to looking out pages about them, and presence in that catalogue is its personal sport, separate from something your weblog does.

Sensible AI search engine optimisation/GEO Tip

When you promote bodily merchandise, test whether or not yours come again with playing cards, presents and service provider names when ChatGPT recommends them. Playing cards imply {the catalogue} is aware of you. A textual content point out with no card means the content material sport is the one one you’re in, and no quantity of web page work modifications what the product name returns.

enterprise is a locations lookup, and it takes a location. On a espresso query, it fired thrice. Listed below are the primary two.

enterprise|Dubai, UAE|specialty espresso Dubai Marina;greatest espresso Dubai Marina
enterprise|Dubai, UAE||Blends Specialty Espresso;Roast Speciality Espresso, Marina;OLAB Espresso Store JBR

The primary name sends search phrases. The second sends precise enterprise names it obtained again from the primary, after which verifies them one after the other with quoted quick searches.

That’s the native model of the shortlist mechanics from the shortlist article, operating on venues as an alternative of manufacturers.

When you run an area enterprise, the unit of optimisation is your entity in the places data.

The lookup returns the identify, the opinions, and the main points, and the online searches solely confirm them.

Ask ChatGPT in your personal class in your personal space and watch the second enterprise name.

In case your venue’s identify seems in it, the lookup is aware of you and the follow-up searches are checking your particulars, so be certain that what’s in your web site matches what the listings say. In case your identify by no means enters that decision, your drawback is the locations information, and your web site can’t repair it.

picture is a picture search. picture|Dubai Marina specialty espresso store|30. It fired alongside the locations lookup.

genui_run calls a widget, with 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":"fairness","market":"USA","locale_override":"en-US"}

A soccer schedule and a inventory chart, requested as features with typed arguments, not as searches in any respect. Extra on what meaning in Thought 5.

Another element from the information seize. The area subject can maintain a listing. theguardian.com;skysports.com appeared on one line, so a search will be scoped to a number of publishers without delay.

Thought 4: The Area Slot

Till this week, when ChatGPT wished to look one particular web site, it wrote the traditional web site:intercom.com in entrance of the question, the identical operator search we SEOs usually use…

The brand new format has a slot for it as an alternative. The final subject on a line is a web site, and that search solely appears to be like there. Checking one firm’s web site is constructed into the device now, sitting subsequent to the question on each line, somewhat than spelled out within the search phrases. Each types nonetheless present up; some strains preserve web site: within the phrases whereas others use the slot.

What this implies in observe hasn’t moved from the shortlist article. The mannequin nonetheless names manufacturers earlier than it fetches something, then goes to their web sites to test what they are saying.

What’s new is how extraordinary that step is. Visiting the web site of an organization it simply named is a typical possibility on each search line, and the mannequin fills the slot itself, from no matter area it already hyperlinks to the model.

If the area in its reminiscence is unsuitable or stale, the probe searches the unsuitable place. I’ve watched that occur as soon as: a probe geared toward a site the model doesn’t run, and the broad search rescued them.

On the question greatest llm visibility device and what does every one price, the mannequin wrote this probe:

web site:profound.ai pricing AI visibility platform Profound pricing 2026

Profound doesn’t run on profound.ai; they run tryprofound.com. The probe returned zero pages from the pool. The model survived anyway. The broad discovery search discovered the proper area; tryprofound.com obtained cited at its pricing web page, and Profound was named six instances within the reply physique. So the unsuitable slot price them one of many mannequin’s 4 searches.

Sensible AI search engine optimisation/GEO Tip

Learn the slot by yourself captures and test the area is one you really run. Rebrands, migrations, and nation domains are the circumstances that produce a unsuitable slot, as a result of the mannequin fills it from reminiscence, and reminiscence lags. A wasted probe price that model nothing on the day I watched it, and I wouldn’t construct a method on the rescue repeating.

Thought 5: Some Solutions Are Changing into Widgets

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

Within the finance reply, the inventory chart confirmed up as a reference named stock_chart pointing at a widget hosted on cdn.platform.openai.com, model 2.0, hydrated from a state payload. Within the espresso reply, the mannequin emitted a genui directive within the reply textual content itself, a map widget with pins wired to the companies the locations lookup returned.

So, for sports activities, shares, locations, and presumably extra I haven’t caught but, a part of the reply is a element OpenAI renders itself. A inventory chart has no hyperlink inside it to win, and the soccer schedule cites no fixtures web page. Complete lessons of question are transferring to surfaces the place the quotation sport, the one each GEO device measures, doesn’t exist.

Sensible AI search engine optimisation/GEO Tip

Run the queries you rank for and notice which come again with a widget the place the reply was. A genui_run line within the seize is the inform. I can’t offer you a tactic for profitable inside a widget, as a result of on what I’ve seen, there’s nothing in a single to win. What you are able to do is know which of your queries have crossed over, and cease counting quotation alternatives on reply varieties that not supply any. In case your visitors is dependent upon being the web page that solutions “what time is the match,” the widget is your competitor now.

It’s the identical name as not optimizing for a keyword whose SERP is an AI Overview and a wall of different options.

The Reddit State of affairs Is Very Very Fascinating

I wish to stroll you thru this one step-by-step, preserving what I can see separate from what I believe it means.

Begin with the road itself, from the stay chat seize.

quick|greatest AI stay chat help Intercom Gorgias Zendesk Ada Tidio Crisp reddit|365|reddit.com

Learn the question. Reddit is known as within the phrases, the area slot is pointed at reddit.com, and the window is 365, the place each model line in the identical name obtained 30. And the six model names are in there with it.

So, what’s occurring right here?

I genuinely had a Joey second. (Pals followers will get it.)

So this search didn’t ask Reddit who the great stay chat instruments are. It requested Reddit what individuals say in regards to the six instruments the mannequin had already picked.

The candidate set got here from reminiscence, and Reddit was fetched for opinion about it.

What got here again is the subsequent tangible reality.

Eighty-four of the 221 entries in that dialog’s retrieval pool had been Reddit threads, greater than from another supply. The dialog had no open-and-read step, so what entered the mannequin’s context was 84 thread snippets, the title and excerpt the search returns, somewhat than 84 threads learn in full.

Then the reply went out with each quotation certain to vendor pages. Zero of the 84 obtained credited, within the one dialog the place I’ve checked citations towards the pool line by line.

The espresso seize then did it once more, tougher.

quick|web site:reddit.com Dubai Marina specialty espresso Dubai|3650|reddit.com

Each mechanisms without delay, web site: within the phrases and the area within the slot, with a 3650-day window. Ten years of threads, on a query the place information websites obtained seven days.

Why Reddit would get a decade on an area question, I truthfully don’t know, however I’ve a couple of ideas on this under. One clue about why a large window suits the query: I requested my AI agent to look Reddit for Dubai Marina espresso threads, and it discovered some as previous as 12 years. That’s the archive a 3,650-day window reaches.

The wire itself can’t verify ages both approach. Virtually not one of the fetched Reddit entries in my seize carried a date, so I can’t let you know how previous the threads ChatGPT really pulled had been.

Promptwatch’s chart from August 18 has Reddit’s share of ChatGPT citations collapsing from about 3.83% to underneath 1% in mid August. Their guess was a supply choice change. A quotation panel solely sees the completed reply. The wire provides the steps earlier than it: the deliberate request, the majority supply, and the zero credit.

Based mostly on what I’m seeing, that is what I believe is occurring. I can’t verify it but, so deal with this as my interpretation of the proof, not a definitive conclusion.

• Content material sitting within the mannequin’s context can form what it writes. (There’s usually a bias.)
• This mannequin fetched opinion about its shortlist instantly earlier than writing opinions about its shortlist, then attributed these opinions to no one.
• I believe Reddit has change into an invisible enter, feeding the verdicts whereas the credit go to vendor pages.
• That will match what I discovered in Half 1, the place information got here from official pages, and verdicts got here from third events, with Reddit chief amongst them. If that division of labour survived and the crediting stopped, the affect continues to be there, and the receipts are gone.
• The opposite potential studying is that the snippets are fetched after which ignored, and the year-wide window is a leftover behavior. Nothing within the visitors separates the 2. The fetching and the zero credit are information I can present you. The affect is my interpretation, and the wire can not show affect with out quotation.

Now let’s unpack this a little bit additional and have a look at the broader context round why OpenAI may need made this variation.

Anti Manipulation And Spam Deterrence

This may very well be a option to cut back the impression of individuals gaming Reddit for AI search engine optimisation/GEO. If ChatGPT is intentionally trying throughout 1 to 10 years of Reddit historical past, latest manipulation turns into a a lot smaller sign.

There are successfully two filters. The candidate pool seems to be chosen first, then Reddit is used to assemble opinions or validate these candidates. That makes merely flooding Reddit with model mentions far much less helpful.

There can also be a deterrence impact. Quotation research present Reddit visibility falling sharply, which makes entrepreneurs suppose Reddit has misplaced its worth. Much less perceived worth means much less spam.

Compliance Or Licensing

For context, OpenAI and Reddit signed an information licensing deal in 2024, so the info entry itself is paid for and public. What that deal says about attribution, I can’t see from the wire. One other risk, and it’s hypothesis, is that OpenAI nonetheless needs the worth of Reddit information with out leaning on Reddit as an apparent, instantly cited supply.

As an alternative of surfacing Reddit pages prominently in solutions, Reddit may very well be used additional upstream as an enter for candidate analysis, opinion gathering, or different inner retrieval processes.

Sensible AI search engine optimisation/GEO Tip

If Reddit has been a part of your AI visibility work, rely the chips your self. Run your queries, and test whether or not the threads you invested in nonetheless get credited on them.

Does this variation make your Reddit campaigns ineffective? Completely not.

When you’ve been doing this for some time, genuinely partaking with Reddit communities and constructing a model, particularly if you happen to survived Reddit’s latest spam purge, keep it up. Two channels survive this variation both approach. Years of threads feed the coaching information that builds the reminiscence shortlist within the first place, and the people studying these threads by no means stopped being patrons. What’s collapsed is the measurable quotation payoff, the chips.

In case your technique was merely spamming Reddit to control AI visibility, although, that is dangerous information for you.

Keep in mind one factor. A thread that will get fetched however by no means cited isn’t a measurable win. Whether or not it nonetheless sways the reply is the open query above.

What Breaks, And What I’d Do

The previous recipe is useless. My final article stated to open DevTools and search the response for queries. That subject not exists. The queries are nonetheless there; they’ve moved into the internet.run tool-call messages because the pipe-delimited strains above. FanoutFox reads the previous subject too, and I’ll be patching it.

The brand new two-minute test. Open ChatGPT in Chrome with DevTools on the Community tab. Ask the industrial query your patrons ask. Filter for dialog, click on the largest response, the dialog payload, not the tiny stream_status and textdocs rows that share the identify. Open its Response tab, press Cmd+F inside that pane, and seek for quick|. The panel-wide search drawer will drown you in JS bundle hits, so search contained in the one response as an alternative. Learn the strains. You’re searching for three issues. Whether or not your model is within the queries, whether or not your area is within the slot, and what window your class runs on.

That test is handbook, and it will get previous on the third reply. FanoutFox, the free Chrome extension I constructed for this collection, reads the identical session information and does the fetched-versus-cited counting for you, per area, per reply. The converts column is the Reddit test from this text as a sorted desk. Its fan-out card reads the previous search_queries subject, in order that card is empty on new conversations till the patch for the pipe format ships, and the remainder of the panel, sources, citations, and the area radar, is unaffected by the change.

That is the third format change I’ve documented since June, and I might put a four-day window on this one solely as a result of I preserve dated snapshots of what the visitors exposes and diff them after each seize session. When you do this sort of work, preserve dated copies of every little thing. What was true on a Sunday was passed by Thursday.

Extra Sources:

This submit was initially printed on Suganthan.

Featured Picture: sdecoret/Shutterstock