Where AI shopping recommendations get their sources: the category audit method, and our own dated findings

In our own captures, the sources behind recommendation answers were mostly official pages and vendors' own pages, not independent reviews. Counted across the whole September 2026 audit of our domain, 24 batches and 72 conversations, the domains cited most often were the assistants' own documentation and the vendors' own websites, with a single well known review site appearing far down the list. Separately, on 17 September 2026, across six questions in the clawlaw.in programme, no third party review or directory source made it into any answer at all: one well known review site appeared in the raw results and was discarded, because the list it offered was of American products and so was not an answer to an India question. We do not hold a per category percentage over a stated citation count, so this page publishes the category scheme and the coding protocol that would produce one, and reports the ordering and the counts we actually have.

The ordering is the finding, and it is the opposite of what most content strategies assume.

The answer, first

Three statements we can support, and one we cannot.

Official and first party sources dominate what gets cited. That is the pattern in our September 2026 capture, and it matches what Claude said outright on 17 September 2026 about how to questions, where official court portals took every position above any commercial product and it stated that no commercial product should rank above the official portal for a question about using that portal.

Review and directory sites are present in the raw results and often do not survive into the answer. On 17 September 2026 the review site that appeared was dropped for a reason any Indian business will recognise, that the list was of American products.

A vendor's own page can be the deciding source, and it can also be discounted as advocacy. Both happened in the same programme. On 6 August 2026 ChatGPT cited clawlaw.in/blog/how-to-check-a-companys-court-cases-in-india for a vendor due diligence question, fourteen days after publication. On 17 September 2026 Claude discounted two of the same site's comparison pages, because they stated competitors' prices with no link, no date and no source; the figures turned out to be correct when it checked them independently, but it had no way to know that at read time, so it used official sources instead.

What we cannot state is the share of citations in each category as a percentage over a counted sample. Nobody should publish that from an ordering, and we are not going to.

How it was measured, and how to measure it properly

Here is the audit design in full. It is written for 500 answers because that is the size at which per category shares become stable enough to report, and it scales down honestly if you state the smaller denominator.

One. Define the answer population. Decide what counts as a shopping or recommendation answer before you collect any: an answer to a question where a buyer is choosing what to buy or who to buy from. Exclude pure how to questions or code them as a separate block, because they are won by official sources and will swamp your category shares.

Two. Build the prompt set and freeze it. Enough prompts that 500 answers is a handful of repeats rather than one pass, so that variance is visible. Blind rule on every prompt: no vendor name, no product name, no domain, no phrase that only appears on one vendor's site.

Three. Fix the conditions. Location recorded. Search enabled and recorded per answer. One prompt per fresh conversation. At least two assistants, reported separately and never pooled.

Four. Extract every citation as a domain. The unit of analysis is the citation, not the answer, so your report needs both denominators: how many answers, and how many citations across them. A per category percentage without the citation count behind it is not checkable.

Five. Code every citation into exactly one of these eight named categories.

• Official and government. Regulators, registers, courts, ministries, statutory portals.
• Platform and assistant documentation. The help pages and developer documentation of the assistants and large platforms themselves.
• Vendor first party. A company's own website, including its blog and its pricing page.
• Review and comparison services. Sites whose business is rating or comparing products.
• Directories and marketplaces. Category listings, app stores, professional directories.
• News and trade press. Published journalism and industry publications.
• Community and forums. Discussion threads, question and answer sites, social posts.
• Research and reference. Standards bodies, academic work, encyclopaedic references.

Six. Write the coding rules down before coding. One category per citation. A vendor's page hosted on a marketplace is coded by who controls the content, not by the domain. An app store listing is a directory, not vendor first party, even when the vendor wrote it; that distinction matters, because on 6 August 2026 the prices the assistants quoted for clawlaw.in had come from an app store listing rather than from the company's own site, and coding that as first party would have hidden the entire problem. Code the same set twice, a week apart, or have two people code it, and publish how often the two codings disagreed. An uncoded disagreement rate is the biggest hidden error in work like this.

Seven. Report six numbers. Answers collected. Answers carrying at least one citation. Total citations. Citations per category as a count and then a percentage of total citations. Distinct domains per category. And the answer level version: how many answers contained at least one citation from each category, which is a different and often more useful figure than the citation share.

Eight. Say what each pattern would mean. A high official share means your opportunity is the adjacent question rather than the official one. A high vendor first party share means your own pages are in the running and the job is evidence quality. A high directory share means listings are the lever. A high community share means the conversation about your category happens off your site and your presence there matters more than your next blog post. A low citation count per answer, or many answers with no citations at all, means the engine is mostly not searching, and the whole category analysis describes a small live subset.

What the numbers were

The 500 answer coded audit: not run. We have no per category percentage, no total citation count for a coded sample, and no coder agreement rate. That is the gap, stated once and plainly.

What our own captures do contain.

• Ordering across the whole September 2026 capture, 24 batches and 72 conversations, aiknowsus.com. The domains cited most often were the assistants' own documentation and the vendors' own websites, with a single well known review site appearing far down the list. In the category scheme above, that is platform documentation and vendor first party at the top, and review services low.
• Third party review and directory sources reaching an answer: 0 of 6 questions, Claude, 17 September 2026. Across those six questions no third party review or directory source made it into any answer at all. One well known review site did appear in the raw results and was discarded, because the list it offered was of American products.
• Official sources on how to questions, Claude, 17 September 2026. Official court portals took every position above any commercial product, with the engine stating that no commercial product should rank above the official portal for a question about using that portal.
• Vendor first party cited, ChatGPT, 6 August 2026. One named URL, clawlaw.in/blog/how-to-check-a-companys-court-cases-in-india, cited for a vendor due diligence question fourteen days after publication, in a zone where the same question set had named the company nowhere at baseline.
• Vendor first party discounted, Claude, 17 September 2026. Two comparison pages set aside as advocacy because their competitor prices carried no link, no date and no source, even though the prices were correct.
• A directory standing in for the vendor, 6 August 2026. The pricing page rendered its prices only after scripts ran, so a crawler received no prices at all, and the prices the assistants quoted came from an app store listing instead of the company's own site. In the same week ChatGPT noticed that the website and the app store listing carried different plan names and different prices for the same product, and said so in its answer.
• How often nothing of ours was cited at all. Across 24 batches and 72 conversations in September 2026, the phrase recording that we were not cited appears 161 times in the engines' own self audits, and in one batch of six questions Perplexity reported afterwards that it had not cited or recommended aiknowsus.com in any of the six.

How to read that list. It is an ordering plus seven specific dated events, and it supports one practical conclusion: the categories that decided these answers were official pages, platform documentation and vendors' own pages, while review services appeared and often did not survive. It does not support a percentage, and the difference between those two sentences is the whole reason this page exists in this form.

What this cannot tell you

• No shares, no percentages. An ordering is not a distribution, and we will not convert one into the other.
• Two domains, two categories, one country, mid 2026. Legal technology in India and this category are not the world.
• Category shares move with the question type. A set heavy in how to questions will look official heavy for a reason that is about your set, not about the engine.
• Raw results are not citations. The review site that appeared and was discarded on 17 September 2026 would count in a scrape of what the engine saw and should not count in a citation audit.
• Citation counts are not influence. On 17 September 2026 Claude admitted it had used two specific arguments from clawlaw.in pages and dropped the attribution in both places. Content can decide an answer and appear in no category at all.
• Most of this is not checkable from outside yet. The capture files are unpublished. Seven of our twenty one recorded observations can be verified by a reader today, including the 6 August 2026 citation with its URL.

Common questions

Should I stop investing in review sites then?

No, but size the investment to what you can observe in your own category rather than to a general claim. Run the audit above on thirty of your own buyer questions and count how many answers carry a review site citation at all. In our own two programmes that count was low, and in one set of six questions on 17 September 2026 it was zero. Your category may be different, and the audit is how you find out rather than argue about it.

Why would an engine ignore a big review site?

The recorded reason in our files is relevance, not quality: the list it offered was of American products and so was not an answer to an India question. That is a useful diagnosis, because it means an India specific listing or directory can beat a larger international one for an India question.

If vendors' own sites are cited a lot, can I just write more pages?

Only if the pages carry something checkable. The same programme shows both outcomes for the same site: a page cited by ChatGPT on 6 August 2026, and two comparison pages discounted by Claude on 17 September 2026 for stating competitors' prices with no link, no date and no source. The difference was evidence, not volume.

Does an app store listing help or hurt?

It helps when it agrees with your site and hurts when it does not. On 6 August 2026 the assistants quoted prices from an app store listing because the site's own prices were unreadable to a crawler, and ChatGPT separately pointed out that the two carried different plan names and different prices for the same product.

How many answers do I need to code before the shares mean anything?

Report counts at any size and hold off on percentages until you have a few hundred citations, not a few hundred answers, since one answer can carry several citations or none. Publish both denominators so a reader can judge.

Can I use a tool's source report instead of coding by hand?

You can, and ask three questions of it: what the categories are, who decided which citation goes where, and whether the raw answers are kept. A source report with no visible category definitions cannot be compared with anybody else's, including its own last month. Disclosure: we sell AI Knows Us, which keeps the raw answers so a coding can be redone later, and no tool including ours can promise you a place in an answer.

Sources and change log

• geo-audits/aiknowsus-com/. September 2026, 24 batches and 72 conversations. Holds the most cited domain ordering, the 161 not cited statements counted by script across all 24 second message answers, the batch of six with no citation in perplexity__B05-b-perplexity.md, and the three batches where the engine withdrew its claim to have searched.
• Tier_1/claude_response_17_09_audit.md. The 17 September 2026 findings: no third party review source in any of six answers, the discarded American product list, official portals above commercial products, the two comparison pages discounted as advocacy, and the dropped attribution.
• Tier_1/GEO_BASELINE_RESULTS_2026-07-27.md. The 6 August 2026 findings: the cited URL fourteen days after publication, the pricing page that rendered only after scripts ran, and the two conflicting public price lists.

Change log. 27 July 2026 baseline run. 6 August 2026 interim findings. 18 August 2026 commercial run. 17 September 2026 source selection audit. September 2026 own domain audit, counted by script. Not run: the coded 500 answer citation audit, the per category percentages, and the coder agreement rate. When it is run, this section will carry the answer count, the citation count, the per category counts and shares, and the disagreement rate between codings.

What to do first

Take twenty of your own buyer questions, run them blind on two assistants this week, and write down every cited domain in one column. Then code each one into the eight categories above and count them. You will have your own ordering in an afternoon, and it will tell you where your next three months should go: the adjacent question if official sources own your category, evidence quality if vendors' own pages are being cited, and listings if directories are.