Source links, brand links and shopping cards in AI answers: how to code each one

To tell a source link, a brand link and a shopping card apart, look at what each one is attached to. A source link is the address an engine gives as the source for a passage it wrote, often a link right after a sentence. A brand link is a link on a company's name, usually in a list of options. A shopping card is a product panel with an item, often a picture or a price. Each supports a different conclusion, and only source links are candidates for citations. Lump them under one label and your report claims more than the answer showed.

In short

• Give every link or card in an answer one label: source link, brand link, product card, place link, video, ad, or unclassified.
• Only source links are candidates for citations. Product cards, place links and similar items are counted, where the engine reports them, as answer features, not as citations.
• A link's visible position and label are evidence about the answer. They are not evidence about whether anyone clicked.
• Engines differ in which of these they expose. A blank means "not reported", never zero.
• Code the answer you have, keep the coding sheet with the run date, and leave the cell unknown when the evidence does not settle it.

Why does the type of link matter?

Because the type decides what you are allowed to conclude. A citation says the engine credited a page for something in its answer. A product card says the answer included a product panel. Neither says the reader clicked, and neither says the engine preferred you.

DiscoveredBy's docs define a citation as one link an AI engine puts in its own answer, pointing at a specific URL as the source for what it just said. A mention is your brand named in an answer. A retrieval is a page the engine read but did not link. These are the key terms the rest of this post builds on.

Add a fourth family and the picture gets more mixed: answer features, the structured parts an engine puts inside its answer, such as shopping, local businesses, ads, video, images and tables. Some links belong to that family, not to the citation family. If you count them all as citations, your citation numbers include things the engine never offered as sources.

What are the link and card types you can see in an answer?

Seven labels cover almost everything. Treat these as the coding scheme; adjust the names to suit your team as long as you keep the distinctions.

Source link. A link the answer gives as the source for a passage. In ChatGPT (app) it can be a small link after a sentence, whose text is the site's address, such as rtings.com. Perplexity marks the sources it used with numbered markers such as [1], and each marker points at a listed source. This is the label that can become a citation.

Brand link. A link to a company's own site, such as a vendor named in a list of options with a link on its name. It tells you the answer pointed the reader toward that company. It is not automatically a source for a claim, and the docs do not say how each engine treats it, so check the engine's rules before you count it as a citation.

Product card. A panel or list item for a specific product, often with a picture or price. The docs call the counted version "shopping".

Place link. A link to a business location, for example a Google Maps place. The docs call the counted version "local businesses".

Video. A link to a video, for example a YouTube video.

Ad. A unit the engine shows as an advertisement.

Unclassified. Anything you cannot place. Keep this label; it protects your other columns.

Which of these does DiscoveredBy actually capture?

It stores source links as citations, and it reports product, local business, video, ad, image and table features as rates on the engines that report them. The docs describe no "brand link" category, and no page describes recording clicks.

Here is what the docs support, taken from the Citations, Answer features and Engines pages.

• Citations. One row per link the engine puts in its answer as a source. A link is attributed to you when its registrable domain matches your project's domain. A subdomain of yours counts as you.
• ChatGPT (app). Citations are the answer's inline sources. Links to a local business's own site, and product and retailer links, are not citations.
• Gemini (app). Citations are the web pages among the answer's sources. Google product links and Google Maps place links are not citations; they are counted as shopping and local businesses among the answer features.
• Google AI Overviews. Shopping counts the products in a shopping block, or the products behind Google's product-viewer links inside the Overview. Video counts when the Overview holds a video part or links a YouTube video.
• Perplexity. Only a source that a numbered marker points at becomes a citation. Other entries in the search-results list are recorded as returned search results, not citations.
• Claude. Per the Citations page, every page its search returned is stored as a citation as well, which the docs cover under retrieved vs cited. Read a Claude citation count with that in mind.

The feature rates are per engine, and they read the structured parts of the response, never the answer text. The docs are explicit that product links, images or cards written into answer text are not counted, with named exceptions: Google's product-viewer links, Google Maps place links among Gemini (app)'s sources, and YouTube videos an AI Overview links.

Two engines that look alike are not: Google AI Mode reports no feature the product counts, and Gemini (app) reports no ads and no video. A blank cell means the engine does not report that feature. It never means zero.

How do you code a link type step by step?

Work down the answer once, top to bottom, and put each link or card through the same four questions. Keep the questions in this order.

1. What is it? Link on a sentence, link on a name in a list, panel with a product, place, video or ad label.
2. What does it claim to support? A passage of the answer (source link), the existence of a company (brand link), or a purchasable item (product card).
3. Where does it point? Your domain, a competitor's domain, a third-party domain, or a search or maps address.
4. What can you confirm? What you can see is the label, the position and the address. Anything about what the reader did next is unknown.

Then pick the label. If a link could be a source link or a brand link, look at where it sits. A link right after a sentence, with the site's address as its text, is a source link. A link on a company name inside a list of options is a brand link. If you cannot tell, code it unclassified and note why.

The manual coding guide (copy this)

Use one row per link or card. Keep a header block for the run itself.

RUN HEADER
Prompt (exact text):
Engine and surface (for example ChatGPT (app)):
Date collected:
Country / city / language of the run:
Collection method (app view or API, per the engines page):
Answer features reported by this engine (from the docs): 
Coder and date coded:

ROW (one per link or card)
#:
Position (nth link or card in the answer):
Visible label or text (as shown):
Address (as shown, tracking parameters removed):
Domain (registrable domain):
Owner: yours / competitor / third party / search or maps address
Type: source link | brand link | product card | place link | video | ad | unclassified
Supports which passage (quote the first 8 words, or "none"):
Counted as a citation by the rules of this engine? yes | no | unsure (see docs)
Counted as an answer feature by the rules of this engine? shopping | local | video | ad | none | unsure
Unknown: (write "click behavior" and any other thing not visible)
Notes:

Every row ends with the same honest line: click behavior is unknown. That is not filler. It stops a reader of the sheet from treating a link as a visit.

A shorter decision table for the type column:

Enlarge image

Enlarge image

Worked example: coding one answer

Illustrative example: Quillstone and its competitors are fictional, and the numbers are made up to show the method.

Quillstone sells document-review software to legal and compliance teams. Its analyst collects one ChatGPT (app) answer for the prompt "best document review software for a compliance team", collected on a fixed date in a fixed country. The answer contains nine links or cards. The analyst codes each row.

The arithmetic, row by row: 4 source links, 2 brand links and 3 product cards make 9 items. Of the 9, 4 are citation candidates (44% of the items, 4 of 9). Of those 4, one points at Quillstone's domain, so 1 of 4 citation candidates (25%) is Quillstone's. Whether a monitoring tool would store all four as citations depends on the engine's rules, so check that separately.

Now see what happens when the analyst does not separate the types. If all 9 items are called "links", Quillstone has 3 of 9 (33%): its citation, its brand link and its product card. That number blends a source, a list entry and a product panel. It looks better than the 1 of 4 the source links show, and it is not the same claim.

The honest report reads: "In this answer Quillstone's domain was cited once out of four sources; Quillstone was also named with a link in the list of options and had one product card out of three. We cannot see whether anyone clicked."

Enlarge image

What this cannot tell you

A coding sheet describes one answer on one date. It cannot tell you these things.

• Click behavior. The answer shows a link. It does not show who used it. To see visits, use a traffic tool and read AI referral traffic as its own measurement.
• Why a link appears. A link's presence is an observation of the answer, not proof of why the engine produced it. Do not write that the engine ranks a source higher because of anything on the page.
• Repeatability. A manual check may differ from a monitoring result, for example when the account, location or wording differs. For the wider reasons, see why your manual ChatGPT check differs from a monitoring result.
• What is not counted. The docs list what the answer-feature counts leave out: brands, prices and merchants inside product cards, and who placed an ad. You can code those by hand, but no count in the product will match them.
• Engines that report nothing. If an engine reports no features, code what you see and record "engine reports none" on the header. Do not turn a blank into a zero.

Common mistakes

• Calling every link a citation. For ChatGPT (app) and Gemini (app), the docs treat product, retailer, business-site and place links separately from citations.
• Reading a source link as an endorsement. A source link says the engine credited a page. Whether the page said good things about you is a separate question; see your website is cited, but your brand is missing from the recommendation.
• Counting a mention as a link. Your brand can be named with no source at all; see your brand is mentioned without a source.
• Merging retrieved and cited pages. A page the engine read but did not link is a retrieval, not a citation. The post on a page retrieved but not cited covers that distinction.
• Comparing shopping rates across engines that do not report shopping. A blank cell is not a low rate.
• Mixing app and API sources in one sheet. Keep them apart, and note the surface on the run header.

Frequently asked questions

Is a link in an AI answer always a citation?

No. A citation is a link the engine gives as the source for what it just said. In ChatGPT (app), per the docs, links to a local business's own site and product and retailer links are not citations. In Gemini (app), Google product links and Google Maps place links are counted as answer features instead.

What is a shopping card in an AI answer?

It is a product panel the engine draws as part of its answer. DiscoveredBy counts it as the "shopping" answer feature on the engines that report it: Google AI Overviews, ChatGPT (app) and Gemini (app). Google AI Mode, Perplexity, Grok, Claude and Gemini (API) show a blank for shopping, which means not reported.

Can I tell from an answer whether anyone clicked a link?

No. The answer shows which links were offered, not which were used. Click behavior needs a separate source, such as your analytics. Treat visible evidence and click behavior as two different observations.

Does a brand link help my citation rate?

Not by itself. Your citation rate is counted from citation rows, which are the links an engine gives as sources. A link on a company name in a list is a different observation, and for ChatGPT (app) the docs say a local business's own site and product and retailer links are not citations.

Why does my Google AI Overview count videos but AI Mode shows none?

Because the two engines report different things. Google AI Overviews counts a video part or a linked YouTube video. Google AI Mode reports no feature the product counts: the docs say its answers arrived as plain answer parts, with images and product cards drawn inside the text, which the product does not read.

Which engines let me compare link types side by side?

Only ones that report the same features. Shopping is reported by Google AI Overviews, ChatGPT (app) and Gemini (app), and web search by the engines that record it. See the Answer features table for the full grid.

Next step

Once you have coded a few answers by hand, you will see which distinctions matter for your prompts. DiscoveredBy keeps citations and answer features as separate views, with the Citations screen for source links and the Answer features card on each prompt's page. The Sources feature covers the citation side, and either view can be set beside your hand-coded sheet. Sign in to DiscoveredBy and check one prompt you already coded by hand.