Skip to content

One ChatGPT answer in eight now cites no source. What it does to your AI visibility report

Goodie's tracking shows ChatGPT answers citing no source rose from 1.5% to 12.2% on 31 October 2026. Which AI visibility metrics break, and what to report.

By Published 14 min read
A row of AI answer cards, each with a link icon beneath it, except one highlighted card that has an empty outline instead.

The short answer

It splits your AI visibility report in two. In Goodie's tracking of 94,851 ChatGPT answers from 9 October to 9 November 2026, answers citing no source rose from 1.5% to 12.2% on 31 October. Citation metrics now cover fewer answers, while brand mentions still cover all of them. Report the no-source rate, citation share and mention share apart.

Key takeaways

  • In Goodie's tracking of 94,851 ChatGPT web answers from 9 October to 9 November 2026, the share citing no source rose from an average of 1.5% to 12.2% on 31 October and stayed above 10% for ten days.
  • Citations per ChatGPT answer in Goodie's panel fell from 5.44 to 4.53 between the October and November 2026 windows, and about 63% of the lost citations came from answers that cited nothing at all.
  • A metric built only on citations covers fewer ChatGPT answers after 31 October, while mention share still covers every answer, so the two belong on separate lines of a report.
  • Trackers start from very different no-source baselines because they run different prompts, tiers and models; compare each tracker with itself before and after the date, never with another tracker.
  • Before treating a fall in citation share as a loss, check whether your share among answers that still cite held steady; if it did, the answer layer changed and your sources did not.

What changed in ChatGPT's citations on 31 October?

The share of ChatGPT answers that cite no source at all jumped in a single day. Goodie, which sells AI search analytics, tracks about 3,000 ChatGPT answers a day through the web interface. In its data from 9 October to 9 November 2026, answers with no cited source averaged 1.5% up to 30 October and 12.2% from 31 October.1 That leaves about one answer in eight with no link for anyone to win.

The break was sharp and it held. Between 9 and 30 October 2026 the daily share never passed 2.1%, and from 31 October it never fell below 10%, across ten days and two weekends.1 Daily answer counts stayed between about 2,800 and 3,100 through the break in late October 2026, so the jump does not come from a smaller or stranger sample.1 Goodie also ran prompts by hand in the ChatGPT interface and saw answers without sources there too.

For a brand, the change lands in the metric most reports lead with. Citation share asks who gets linked when an answer links anyone. When one answer in eight links nobody, that share describes fewer of the answers your buyers read, even if nothing about your site changed.

ChatGPT web answers, 9 October–9 November 2026

One day, a different answer layer

1.5%of answers cited no source, 9–30 OctGoodie, Oct 2026 1
12.2%of answers cited no source, 31 Oct–9 NovGoodie, Oct–Nov 2026 1
4.53citations per answer after the break, down from 5.44Goodie, Oct–Nov 2026 1
Source: Goodie, zero-source answers study, 94,851 ChatGPT web-interface answers from its own tracking.

A disclosure before the detail

Sigzen AI sells AI visibility audits and programmes at published prices, so we have an interest in how this gets measured.7 Goodie sells in the same market and says as much in its own study. The method below works with any tracker that stores raw answers, and most of it is a spreadsheet exercise.

What Goodie measured, and what it cannot tell you yet

Goodie measured how many answers carried citations and how many citations each answer carried. Its baseline window, 9 to 30 October 2026, held 65,431 answers, 984 of them with no source.1 The window from 31 October to 9 November 2026 held 29,420 answers, 3,601 of them with no source.1 Citations per answer fell from 5.44 to 4.53 between the two windows, a drop of 17%.1

Most of the loss came from answers that went silent. Goodie attributes about 63% of the citations lost after 31 October 2026 to answers that cited nothing, and 37% to answers that still cited, slightly less, at 5.16 sources each against 5.52 before.1 Put simply, roughly one answer in nine lost every source and the rest lost very little.

What the study does not say

It does not say why. Goodie points to an OpenAI release note of 29 October that made GPT-6 answers more concise for Free, Go and logged-out users. It adds that the note mentions neither search nor citations, and that OpenAI has not confirmed a link.1 The study does not name the model or login state behind its collector, the countries it covers or the kinds of prompts it tracks. Goodie has not yet split the change by prompt type and says it needs more days of data.

The weeks around the break

  1. GPT-6 Sol reaches paid tiers on 7 October and GPT-6 Luna reaches Free and Go on 8 October.2

  2. Goodie's baseline window opens.1

  3. OpenAI release note: more concise answers for Free, Go and logged-out users, as reported by Goodie.1

  4. The share of answers with no source jumps and stays up.1

  5. OpenAI launches Business Hub for verified businesses, as reported by Goodie.1

  6. Goodie publishes the study.1

Sources: OpenAI release notes; Goodie.

Why do other trackers report a far higher no-source rate?

Each tracker measures a different ChatGPT and a different set of questions. Goodie's baseline of 1.5% no-source answers in October 2026 sits far below other published figures.1 Ahrefs, which also sells AI visibility tools, found that only 14.4% of 66.6 million free, logged-out ChatGPT responses on GPT-5.6 Luna triggered a web search in August and September 2026.3 Most of those responses could not have carried a citation at all.

Similarweb's US data points the same way: about 6.8% of ChatGPT answers carried a citation in May 2026, up from about 1.6% in June 2025.4 Those samples include small talk, coding help and every other use of ChatGPT. A brand tracker runs commercial questions that usually send the model to the web, so its no-source rate starts low. Goodie notes that another tracker's US panel started from a very different level before an earlier change in March, and that only the direction compares.1

Our view: never compare one tracker's no-source rate with another's. Compare each tracker with itself, before and after the date, and treat agreement in direction as the signal. Two panels can start at very different levels and still record the same product change.

Which AI visibility metrics break when answers stop citing?

Citation metrics break first, mention metrics hold, and blended scores drift. Most reports carry one of two citation measures. The first is the share of tracked answers that cite your site. It falls mechanically when answers stop citing anyone, even if you lost nothing. The second is your share of all citations, and it can rise or fall depending on which answers went silent.

Mention share is sturdier. A no-source answer still names brands, so a metric that counts names in the answer text keeps covering every answer. That is why our method for coding how answers describe a brand reads the answer, not only the links. Referral clicks from ChatGPT can dip as well, since fewer answers carry a link to click.

Blended scores are the hardest to read. A score that mixes citations and mentions with fixed weights will move after 31 October whatever you did that week. Our view: split any blended score into its parts before you try to explain a change in it, and keep the parts on the report.

Which metrics a no-source answer affects

Built on citations

  • Share of answers citing your site
  • Your share of all citations
  • Owned versus third-party citation mix
  • Referral clicks from answers

Built on the answer text

  • Share of answers naming your brand
  • How strongly you are recommended
  • Whether the facts stated are right
  • Where you sit in a list of names
Illustrative. The left column moves when answers stop citing; the right column still covers every answer.

How to check your own data for the 31 October break

Put a marker on 31 October 2026 and compare equal windows either side. Goodie compared the three weeks from 9 to 30 October with the ten days that followed. You can run the same comparison with any tracker that stores raw answers, since most exports let you count answers with an empty source list. The test takes an afternoon.

Testing your ChatGPT data for the break

  1. Mark the dates

    7 and 8 October for GPT-6, 29 October for the release note, 31 October for the break.

  2. Count silent answers

    Per day and per engine, the share of tracked answers with an empty source list.

  3. Recompute on cited answers

    Your citation share among answers that cite anyone, before and after the date.

  4. Check mention share

    The share of all answers that name you, which should not jump on the date.

  5. Read a sample by hand

    Open some silent answers to confirm no link was missed by the parser.

Illustrative. The order we use; any tracker that exports raw answers can run it.

Then read the result. If your share among cited answers held steady, your sources are fine and the drop sits in the answer layer. If it fell, something in retrieval moved as well. Our breakdown of the August Reddit shock shows what a retrieval-layer shift looks like in the same kind of data.

Keep the engines apart. The Goodie figures cover ChatGPT only. Goodie says its team saw Google's AI Mode cite fewer sources around 5 November, but it published no figures for that. Check each engine you track on its own dates before you write one story about "AI" in the report.

A figure studies a report chart in which one bar is only a dotted outline, standing for answers that cite nothing.
Most answers still reach out to sources; a growing few now stop short.

What wins an answer that searches nothing?

Memory wins it, and memory is built slowly and mostly outside your own site. Goodie describes two layers in ChatGPT: a search tool that picks sources, and a layer above it that decides whether to search at all and how to show the result. It places the 31 October change in the upper layer, and the August Reddit change in the search tool.1 An answer that never searches is written from what the model already holds.

What the model holds comes from training data. That means reviews, lists, forum threads, press coverage and your own pages as they stood before the cut-off. Our guide to splitting a prompt set into memory and retrieval prompts shows how to tell the two apart in your own runs. The levers for memory prompts are consistent third-party presence and time, and they move when a new model ships.

Our view: if a share of your buyer prompts now goes unsourced, the memory score deserves its own line and its own owner. It measures work your PR, partner and review teams already do, against the questions buyers actually put to an AI engine.

Pages still decide most cited answers

Retrieval still governs the large majority of the answers Goodie tracks. Goodie's own advice puts "be retrieved" first, because search still decided close to nine in ten of the answers it tracked from 31 October to 9 November 2026.1 Its second and third pieces of advice, "be remembered" and "be registered", cover the silent answers and the new business profile.

The answers that keep sources also read widely. Semrush's 2026 AI Visibility Index, covering 126 million US prompts from January to April 2026, found ChatGPT cited an average of 15 sources per response, against 3 for Gemini.5 Across 11.84 billion citations on eight engines from 16 April to 16 July 2026, Profound found about 57% went to company-owned sites, and 47% on ChatGPT.6 Semrush counts sources differently from Goodie, so keep the two figures apart.

The page work does not change. Keep prices, features, coverage and company facts in plain text on short, well-titled pages, so a searching answer can read them. Our AI visibility service starts with a map of which buyer prompts search, which of your pages they reach and which return nothing.

Where ChatGPT's lost citations went after 31 October 2026

Answers that cited nothing at all63%
Answers that still cited, slightly less37%
Source: Goodie, share of lost citations, 9–30 October vs 31 October–9 November 2026.

Does OpenAI's Business Hub replace your own tracking?

Business Hub adds a view from inside ChatGPT, but it does not replace a prompt set you control. According to Goodie, OpenAI launched it on 10 November 2026 within a beta for businesses in ChatGPT, as its first analytics product for brands.1 Goodie reports that it shows verified businesses their mentions, profile views, website clicks and discovery topics, while the prompts stay hidden.

That makes the profile worth claiming. Goodie's "be registered" advice is to claim it, and a verified profile with correct facts costs little to keep. It also lets a business hand OpenAI its details and product feed directly. It does not answer the questions a report has to answer: which buyer questions you lose, to whom, on which engine, and whether a change on your side moved anything.

Our view: claim the profile, check every fact in it, and keep your own prompt set as the measure. A dashboard an engine publishes about itself is useful context, not an audit. Our before-and-after method for model updates still applies: freeze the prompts, compare equal windows and report with intervals.

Explaining the drop to leadership

Lead with the cause, then the evidence, then the plan. A chart that falls in the first week of November will draw questions, and a vague answer invites a budget cut. Say that ChatGPT began citing no source in about one tracked answer in eight on 31 October 2026, in one vendor's panel, and that the change sat above the search layer.1 Then show your own split: the share among cited answers and the mention share across every answer.

Be honest about what nobody knows yet. OpenAI has not confirmed a cause, the Goodie panel covers ten days after the break, and an earlier rise in silent answers in March faded within about two months, according to Goodie.1 The change may reverse, settle or spread to other engines. A report that says so keeps its credibility when the next release moves the line again.

Our view: put the decision in front of leadership rather than the drop. The decision is whether the memory line gets an owner and a budget of its own, since that is the part of ChatGPT your pages cannot reach on their own. The retrieval work continues as planned either way.

What your report should show from November

Show three lines for ChatGPT instead of one, each dated. The first is the no-source rate: the share of your tracked answers that carry no citation, per engine and per week. The second is citation share among answers that cite anyone, which separates your sources from the answer layer. The third is mention share across every answer, the line a silent answer cannot hide.

Annotate the chart. Mark 7 and 8 October for GPT-6, then 29 and 31 October for the release note and the break.2 Goodie recommends the same: mark platform changes on your charts and track search rate separately from citation share. A report without those markers invites a team to claim credit for a product change, or take the blame for one.

Repeat the runs. Our audit runs 150 buyer prompts across ChatGPT, Gemini, Microsoft Copilot, Google AI Overviews and Google AI Mode, each on three separate days, and reports with confidence intervals.7 Re-run the split of memory and retrieval prompts after each release, since a prompt can move between groups when the answer layer changes.

Our view: the 31 October break is a measurement lesson more than a visibility loss. Teams that report three lines will see it for what it is within a week. Teams that report one blended score will spend November explaining a drop they did not cause.

Sources

  1. Goodie, ChatGPT's zero-source answers jumped 8x overnight: 94,851 ChatGPT web-interface answers from Goodie's own tracking, 9 Oct–9 Nov 2026; geography not stated (Nov 2026)
  2. OpenAI, ChatGPT release notes: GPT-6 Sol for paid tiers from 7 Oct 2026 and GPT-6 Luna for Free and Go from 8 Oct 2026 (Oct 2026)
  3. Ahrefs, ChatGPT query fan-out study: 66.6M free, logged-out GPT-5.6 Luna responses, Aug–Sep 2026, all countries; factual subset 20,000 US prompts; citation ranks US, Mar–Sep 2026 (Oct 2026)
  4. Similarweb, AI search stats 2026: ChatGPT answers with a citation, US, Jun 2025–May 2026 (Jul 2026)
  5. Semrush, 2026 AI Visibility Index: 126M US prompts, Jan–Apr 2026 (Jun 2026)
  6. Profound, where AI citations come from: 11.84B citations, 8 engines, 16 Apr–16 Jul 2026 (Jul 2026)
  7. Sigzen AI, published prices (Nov 2026)

Questions readers ask

  • A zero-source answer in ChatGPT is a reply that carries no citation or source link at all. It is written from what the model already holds or from a search it chose not to show. Brands can still be named in it, so it counts for mention tracking, but it offers no link to win, no referral click and nothing for a citation report to measure.

Keep reading

Free. Enter your domain and Sigzen AI checks how five engines answer your buyers' questions, against up to three competitors.

No account needed. Two minutes to submit; we email the full five-engine report when the check has run.