Chapter 6. Corrections: What Our Earlier Posts Got Wrong, and What Changed

AEOsome Research. Vijaya Kumar Channalli. Chapter 6 of the AI Shopping Visibility Study 2026, version 2026-10. The hub carries the abstract, the ten headline findings, the methods summary, and the other chapters. This chapter is the public record of what the re-analysis changed.

Abstract

Between April and October 2026 AEOsome published nine articles drawing on the two datasets this study re-analyses. Before the study was written, every number in those articles was listed in a claims register and recomputed from the raw data under one set of definitions. Of 31 registered claims, 20 reproduced exactly or within rounding, three reproduced in direction but not in level, and eight were wrong or misleading enough to need a public correction. This chapter states each of the eight: what was published, what the data says, why the original was wrong, and how serious the error is. The two largest are a citation-type classifier that typed 99% of cited domains as retail, which produced the published claim that "retail websites account for 89 to 99% of citations", and an intent table labelled de-duplicated that contained seven raw-basis values. The chapter closes with the versioning rule now applied to every number on the site, and the list of posts updated.

1. How to Read This Chapter

Each correction has four parts. Published quotes the claim as it appeared. Corrected gives the value or statement that replaces it. Why explains the error. Severity is High when a headline claim in a live article is wrong in substance, Medium when the direction holds but the level or the generality does not, and Low when the number drifts but the conclusion stands. Section 11 lists every live article touched, with the corrections applied to each.

All corrections were made in October 2026 and the articles carry a dated note. The original wording is preserved in the site's version history and in the vault that holds the datasets. Nothing was removed silently.

Table 1. Summary of the eight corrections.

2. Correction 1. Retail Does Not Account for 89 to 99% of Citations

Published. "Retail websites account for 89 to 99% of all citations across every intent stage; brand-owned sites peak at 6.1%." In the 8,520-query article, the product-description-structure article, and the diagnostic guide.

Corrected. Re-typed with one rule set, Study 1's cited domains are 46% Editorial, 27% Retail, 10% Brand-owned, 7% Marketplace, and 3% Community. By stage, Retail runs from 8.6% at Post Purchase to 48.1% at Purchase Execution; Editorial from 23.6% to 75.1%; Brand-owned from 2.0% to 18.0%, peaking at Attribute Constrained.

Figure 1. The published retail share by stage against the re-typed shares. The published line is the output of a classifier that assigned Retail to almost every domain.

Table 2. Citation types by stage, Study 1: published against corrected.

Why. The engine's domain-type field labelled 14,898 of Study 1's 15,034 cited domains as Retail, including vogue.com, glamour.com, healthline.com, mayoclinic.org, and wikipedia.org. The published analysis took that field at face value. The study replaced it with a single rule set applied to both datasets, documented in chapter 7, and reviewed the 300 most-cited domains by hand.

What survives. Retail is still the leading type at the two purchase-end stages, and brand-owned pages still do best at the attribute stage. The corollary that "brand sites barely get cited" does not survive: brand domains are one citation in ten overall and nearly one in five at Attribute Constrained.

Severity. High. Three live articles and one concept page carried the figure.

3. Correction 2. The Study 2 Intent Table Mixed Raw and De-duplicated Values

Published. A table of card rate by intent stage for Study 2, labelled "22,335 de-duplicated responses", in the "Is AEO worth it" article and on the intent-framework concept page, with Attribute Constrained at 50.2, Purchase Execution 45.1, Scenario Confirmation 42.6, Comparison 37.4, Category Exploration 37.2, Validation 27.9, and Problem Recognition 23.2. All values in this section are percentages.

Corrected. Those seven values are computed on the 27,255 collected answers including retries, not on the 22,335 de-duplicated ones. The de-duplicated values are Attribute Constrained 47.0, Purchase Execution 43.0, Scenario Confirmation 39.6, Comparison 35.7, Category Exploration 37.8, Validation 26.2, and Problem Recognition 20.4. Budget Framing (54.8) and Post Purchase (14.7) were already correct.

Figure 2. Card rate by stage, as printed against corrected. The raw basis over-states the middle stages by two to three points because retries happened disproportionately on card-bearing questions.

Table 3. Card rate by stage, Study 2: as printed against corrected. Values are percentages.

Why. The de-duplication was applied to the overall and platform rates on October 2 and the corrected figures were recorded, but the intent table on the concept page was updated from the wrong column of the summary file. The vault's own summary lists the seven raw values under a heading that says they are never to be published. The guard that now checks every draft against that list was added after this error and has since blocked two further drafts.

Severity. High. The error is small in magnitude but the table was labelled with a basis it did not have, and the study's own rule is that every Study 2 rate is on the de-duplicated basis.

4. Correction 3. ChatGPT's Card Rate Is Not One Number

Published. "ChatGPT shows product cards in 28.5% of answers", as a property of the platform, in the 22,335-query article, the "Is AEO worth it" article, the diagnostic guide, and the platform concept page.

Corrected. 28.5% is the mixture of three served models. The small model, gpt-5-3-mini, answered 60% of questions and showed cards 3.7% of the time; gpt-5-5 answered 30% at 65.3%; gpt-5-3 answered 9% at 72.2%. On the full models ChatGPT's card rate is above Google AI Mode's. Every ChatGPT figure in this study is reported by served model.

Figure 3. The published rate and the three rates behind it.

Figure 4. ChatGPT card rate by stage and served model. The full models exceed 90% at Budget Framing and Attribute Constrained; the small model does not reach 13% at any stage.

Table 4. ChatGPT by served model. Overall, and by stage with the number of answers in brackets.

Why. The collection recorded the served model with each answer but the published analysis did not split on it. The 28.5% figure is correct for the May 2026 collection and will not replicate, because the next run's rate depends on the share of questions routed to the small model that day.

Severity. High. The number is reproduced in four places and is the one most likely to be quoted.

5. Correction 4. Rating Earns Position, Not Presence

Published. "Average product rating predicted carousel rank better than brand recognition; brands above 4.70 outranked names near 4.20." In the reviews article, the ranking-signal-transfer article, the diagnostic guide, and the product-review-signals concept page.

Corrected. Half of the claim holds. Inside a carousel, rating predicts position: the within-carousel correlation between rating and rank is negative (mean Spearman of -0.21 over 3,821 carousels), and one standard deviation of rating doubles the odds of rank one (chapter 4). Across brands, rating does not predict presence: over 110 brands with 50 or more slots, the correlation between a brand's average rating and the number of carousels it reaches is -0.24. Quince at 4.74 stars is in 13% of carousels; Lulus at 4.30 is in 30%.

Figure 5. Average rating against carousels reached for five dress brands. The two highest-rated brands reach the fewest carousels.

Table 5. Rating against rank and against presence.

Why. The published comparison was read off a list of brands ranked by average rating. There is no brand-recognition variable in either dataset, so the claim that rating beats recognition had no measurement behind it; it was an inference from the order of the list.

Severity. Medium. The articles are reframed: rating moves a product toward the front of a shelf it is already on, and does not get it onto the shelf.

6. Correction 5. Naming a Brand Does Not Suppress the Shelf Everywhere

Published. "Naming a brand reduces the carousel rate to 21.1% from 66.6%", in the 8,520-query article, generalized as a rule of AI shopping.

Corrected. True for ChatGPT in Study 1. In Study 2 brand-named questions produced cards more often than generic ones on Google AI Mode (70.2% against 62.4%), Copilot (47.8% against 43.4%), and ChatGPT (31.1% against 27.6%); only Perplexity went the other way. The Study 1 effect is mostly stage mix: brand names cluster in Comparison and Validation questions, which have low card rates on every engine.

Figure 6. Card rate for brand-named against generic questions, Study 1 and Study 2.

Table 6. Brand-named against generic questions, by slice.

Why. One engine, one category, one month was generalized to the medium. What does hold across both studies is the slot share: a named brand shares its shelf with other brands a third to half of the time (chapter 5).

Severity. Medium.

7. Correction 6. The Conversational Collapse Is a February 2026 ChatGPT Result

Published. "Conversational phrasing collapses an otherwise saturated stage from 100% to 25.8%", in the product-description-structure article and on the intent-framework concept page, presented as a property of AI shopping.

Corrected. The result reproduces exactly in Study 1: at Attribute Constrained, Short, Medium, Long, and Messy phrasings trigger a carousel 98.7% to 100% of the time and Conversational 25.8%. In Study 2 the same stage on ChatGPT's full models runs from 90.6% to 98.5% across styles with Conversational at 94.8%; on Google AI Mode the spread is 13.8 points and on Copilot 2.9. Across all stages, the largest style effect on any Study 2 engine is 10.7 points (Google AI Mode) and ChatGPT's is 4.8.

Figure 7. Attribute Constrained card rate by phrasing style: Study 1 against three Study 2 engines. The collapse exists in one bar group.

Table 7. Style effects: the attribute stage, and all stages by engine.

Why. The February ChatGPT models treated a conversationally phrased attribute question as a chat turn rather than a shopping request; the May models and the other engines do not. The finding was real and specific, and it was published as general.

Severity. Medium. The article now states the engine and the month.

8. Correction 7. Signal Rates Were Computed on the Wrong Base

Published. In the 22,335-query article and the concept page: deal-seeking questions get cards 48% of the time against 35% without; price 47 against 34; problem 38 falling to 21; comparison 36 against 32; urgency 36 against 31.

Corrected. The signal tags exist for six of the seven categories; Electronics has none, which the article did not say. On the 18,736 de-duplicated answers that carry tags: deal 42.0% against 32.2%; price 45.6% against 32.0%; problem 19.8% against 35.6%; comparison 30.0% against 33.2%; urgency 32.4% against 32.8%. Four of five directions hold. Urgency has no effect.

Figure 8. Signal present against absent, published against corrected. Urgency's published four-point gap disappears.

Table 8. Query signals, published against corrected.

Why. The published rates were computed on the collected rows including retries and did not restrict to the categories that carry tags. The urgency result in particular was an artifact of retries.

Severity. Medium.

9. Correction 8. Brand Counts and Diversity Figures Were Inflated by the Brand Field

Published. "Comparison questions collapse brand diversity to 3.28 unique brands" (9-stages article); "2.81 against 7.18 brands when a brand is named" (8,520-query article); per-brand carousel counts such as Lulus 1,515 and Petal & Pup 1,208 (9-stages article, reused from an archived post); "Attribute Constrained shows the highest brand diversity at 7.36" (concept page).

Corrected. The engine's brand field is truncated to the first word, so "Old Navy" becomes "Old" and "Petal & Pup" splits from "Petal and Pup". With brands resolved from product URLs and titles (chapter 7): Comparison 2.07 brands; brand-named 1.64 against 5.10; the most diverse stage is Category Exploration (5.5) and Attribute Constrained is fourth (4.8). Per-brand counts move by 2 to 12%: Lulus 1,528, Petal & Pup 1,071, Old Navy 790, Fashion Nova 714, Quince 662.

Figure 9. Carousel counts for six brands, published against resolved.

Table 9. Brand counts and diversity, published against corrected.

Why. A field that looked like a brand name was used as one. Every direction in the published claims holds; every level is 35 to 40% lower.

Severity. Low.

10. Smaller Clarifications

1. Intent against platform. The intent-framework concept page said both that "intent is a stronger predictor of product citation than platform choice" and, lower down, that "platform rates are the largest single effect". The second is right: platform explains a 59-point spread, intent 40 points, category 22, and platform alone predicts a card better than every other feature combined (chapter 4). The first sentence was removed.
2. Home and Toys have no Budget Framing or Post Purchase questions. Their category rates are computed on a different stage mix from the other five categories. The category chapter says so; the earlier articles did not.
3. The Electronics deep dive (an internal analysis, not a live article) was computed on 2,129 collected rows including retries. On the 900 de-duplicated rows the card rate is 40.8% rather than 40.2%, Apple moves from fourth to second among brands, and Best Buy holds 27% of merchant links rather than 32%. Conclusions unchanged.
4. "Deduplicated" in earlier articles. Where an earlier article uses the word, the number was checked against the summary file and is correct unless listed in this chapter.

11. The Versioning Rule

From this release, every number AEOsome publishes from these datasets carries three things: the study it comes from (Study 1 or Study 2), the basis (for Study 2, always the 22,335 de-duplicated answers; the 27,255 collected answers appear only with the words "collected" and "retries removed"), and, for any ChatGPT figure, the served model. Figures are computed from the published scripts, not copied between articles. When an article is updated, it gains a dated note at the point of change and a line at the end linking to this chapter. When the datasets are re-collected, the new numbers get a new version label and the old ones stay citable under the old one.

12. Articles Updated

Table 10. Live articles updated with this release. Each carries a dated correction note.

Each updated article links to the chapter that carries the corrected figure. The earlier wording is retained in the site's revision history.

13. Limitations of the Corrections

1. The corrections are to numbers AEOsome computed. They do not re-run the collection, and a new collection would produce different rates again, for the reasons in chapter 5.
2. The re-typed citation shares depend on the rule set in chapter 7. A different rule set would move the Editorial and Retail boundaries by a few points; it would not restore a 89 to 99% retail share.
3. The brand resolver recovers brands from product URLs and titles and leaves about 9% of Study 1 slots and 5% of Study 2 slots unresolved. Per-brand counts have that uncertainty.
4. Claims that were qualitative (the seven-gate diagnostic sequence) are discussed in the hub's discussion section, not corrected here, because the data can neither confirm nor refute most of them.

Navigation. Previous: Chapter 5. Ten surprises · Next: Chapter 7. Methodology and data · AI Shopping Visibility Study 2026 (hub, abstract, and all chapters).

About the author. Vijaya Kumar Channalli is a serial SaaS founder and the founder of AEOsome (Tartu, Estonia), which helps eCommerce brands get their products recommended when shoppers ask AI what to buy. Read how we write, source and correct posts in our editorial policy.

Sources: AEOsome Research, AI Shopping Visibility Study 2026, chapter 6. Study 1: 8,520 ChatGPT answers, women's dresses, United States, February 28 to March 1, 2026. Study 2: 22,335 de-duplicated answers across ChatGPT, Google AI Mode, Microsoft Copilot, and Perplexity, seven categories, United States, May 5 to May 14, 2026. Hub: https://www.aeosome.com/blog/research/ai-shopping-visibility-study-2026/