AI Search Tool Rank
All posts
By AI Search Tool Rank TeammethodologyChatGPTAPIMCP

AI answers with widgets: what text-only tracking misses

ChatGPT can answer with live election results boxes, not just text. Why text-only tracking misses what readers saw, and how to pull answer widgets from the API.

Most AI visibility tracking stores an answer as a block of text. For a long time that was close enough, because answers were mostly text. Ask ChatGPT about a race on election night, though, and the answer can include a results box: the Associated Press tally with vote totals and whether the race has been called, a box showing the Senate balance of power, cards with each candidate's photo, name, and party. The written part sits around those boxes. A tracker that saved only the written part saved a different answer from the one the reader looked at.

Election questions are not a niche use of these tools either. The Washington Post's analysis of October 24, 2026, which used Promptwatch data, sent a voter-guide request to six chatbots for an address in every U.S. county and analyzed more than 3,000 responses. When answers on a topic that sensitive get studied at that scale, the boxes are part of what has to be checked.

This post covers what goes missing in a text-only record, why the box and the sentence next to it can disagree, and how a data team can pull the widgets out as structured records.

An answer is what the reader saw

Think about how a person reads an election answer. Their eye goes to the box first, because that is where the numbers are. The paragraph below may summarize, add context, or quote figures of its own. If you are auditing what ChatGPT told people, the box is part of the claim, and in many cases it is the main part.

A text-only record loses that content outright. Vote totals, call status, a candidate card: none of it has to be repeated in the prose. Search a text archive for a candidate's name and you can miss every answer where the name appeared only on a card.

It also loses placement. A box at the top of an answer and the same box under three paragraphs are read differently, and you can't tell which one you had from the text alone. Promptwatch's response view shows each box in the place ChatGPT put it, and its API returns the widgets in the order they appeared.

Why the box and the text can disagree

The results in a box are live as of the moment the answer was collected. The figures in the text are whatever ChatGPT chose to write. Promptwatch's tooltip on these boxes says it directly: because the results are live from when the answer was collected, they can differ from the figures ChatGPT quotes in its text.

On a night when counts move by the minute, that gap is worth recording rather than smoothing over. An answer whose sentence says one total while its box shows another is exactly the kind of response an auditor wants to find. You can only find it if both halves were stored.

There is a second, smaller source of difference. The boxes in Promptwatch's response view are rebuilt from ChatGPT's data, and each one carries a label saying so, with a tooltip explaining why it can look different from what ChatGPT drew. The boxes only show data from ChatGPT's answer. Nothing is filled in from another source, so a rebuilt box tells you what ChatGPT sent, not what the final result turned out to be.

Candidate cards are not recommendations

ChatGPT's candidate cards show a photo, a name, and a party. If your tooling treats every named entity in an answer as a mention, a candidate card reads like the assistant putting someone forward. It isn't doing that. The card identifies a candidate, and that is all it does.

Promptwatch labels these as candidate cards instead of recommendations. For anyone reporting on how evenly an assistant treats candidates, that label is the difference between a fair finding and a misleading one. A count of "times ChatGPT surfaced candidate X" built from card appearances would mostly measure how often a card was shown.

Your visibility scores do not change

Promptwatch states that its scores are unchanged by these boxes. That is the right call for anyone tracking a brand over time. Visibility Score rates each response on placement, share of attention, and depth, and Share of Voice counts brand mentions. If an AP results box suddenly started counting toward either, a trend line that spans the change would jump for reasons that have nothing to do with what the assistant thinks of your brand.

So the widgets add evidence without moving the metrics. You can look at them when a response needs a closer read, and your dashboards keep measuring the same thing they measured last month.

Getting widgets out as data

The response view is fine for reading one answer. A data team needs records. On November 6, two days after the response view change, Promptwatch started returning the widgets through its REST API and MCP server, exactly as the provider sent them.

  • GET /v2/responses/:id and the MCP getResponse tool return a components list with every supported widget in the answer, in order.
  • GET /v2/responses and the MCP listResponses tool return the same list on each response when you pass includeComponents=true.
  • Each component has a type, the model that showed it, and its raw payload as received.

The REST API lives at https://server.promptwatch.com/api/v2 and takes your key in an X-API-Key header, so a list request with widgets looks like this:

GET https://server.promptwatch.com/api/v2/responses?includeComponents=true
X-API-Key: <your key>

The hosted MCP server at https://server.promptwatch.com/mcp exposes the same data to Claude, ChatGPT, Cursor, and other MCP clients, which is handy when an analyst wants to ask questions of the archive without writing a script.

Be precise about scope. The supported widgets to start are ChatGPT's election results per race, the Senate balance of power, and voting information. That is the whole list. The dashboard's response view now also shows ChatGPT's voting information box, which it used to leave out. Build on what is documented, read the type field on every component, and have your code skip types it doesn't recognize instead of guessing at them.

Treat the payload as raw, because it is. Its shape is whatever ChatGPT sent, and ChatGPT does change its formats: Promptwatch had to re-read Shopping answers after ChatGPT switched to a new product card format around October 15. Store the payload as received and parse it in a separate step, so a format change breaks a parser you can fix and not the archive itself.

What a data team can do with it

Four uses come up quickly once the components are in a table.

Audit what users saw

Join each response's text with its components and flag answers where the figures in the prose and the figures in the box disagree. That gives an editor or a compliance reviewer a short list to read by hand, not every response.

Archive answers around an event

Pull responses with includeComponents=true across the window you care about. The REST API and MCP accept date ranges of up to 92 days per request. Because each box holds results as of collection time, the archive becomes a record of what was on screen when each answer was captured, which is what you want when someone asks later what ChatGPT was showing at the time.

Count how often widgets appear

Every component names the model that showed it, so you can measure how many tracked prompts trigger a box and on which model, separate from how often the text mentions anyone.

Collect more often when it matters

A daily run will miss most of what happens on a results night. Monitors can track every 12, 6, or 1 hours on Business, Agency Scale, and Enterprise, and since November 4 each of those runs lands inside its own time slot. Pair a faster monitor with the components list and you get a timeline of boxes, not a single snapshot.

Keep the API rate limits in mind when you plan a backfill. Essential allows 500 requests an hour, Professional 2,000, and Business 10,000 (250 a minute). Enterprise allows 24,000 an hour and 400 a minute. On agency plans it is 500 an hour on Kick-off, 2,000 on Growth, and 10,000 on Scale.

Beyond election boxes

The underlying problem is bigger than elections: an AI answer is whatever the product rendered, and text is only part of it. Promptwatch captures some other non-text surfaces through separate features. ChatGPT Shopping covers ChatGPT shopping carousels and Amazon Alexa product results (Business and Enterprise). Ads Radar captures sponsored ads inside ChatGPT Search, Google AI Overviews, and Google AI Mode answers (Business, every agency plan, Enterprise). Those are their own reports. They are not part of the components list, which for now covers only the election and voting widgets above.

Where Promptwatch fits

None of this works if the tracker never sees the rendered answer. Promptwatch monitors ChatGPT Search through its real interface, so the response it records is the one the product showed, and the supported boxes come along with the text. The labels for rebuilt boxes and candidate cards keep a reviewer from reading too much into them, the scores stay stable, and the components list hands the raw widget data to whoever needs to analyze it. MCP and API access is included on every paid plan.

If your team reports on what AI assistants told people during an event, and not only on whether your brand was named, that combination is the deciding factor. A tracker that stores only the text and a mention flag loses the box entirely.

Next step

If you want to see the components list on your own prompts before wiring it into a pipeline, book a demo with Promptwatch and ask to walk through a response with a widget in both the response view and the API.