How do AI Overviews and ChatGPT pick which coupon sites to cite?
Google and ChatGPT filter coupon sites for authority and quality, but neither checks if a Black Friday code actually works at checkout.
Google AI Overviews and ChatGPT both run candidate coupon pages through authority and quality filters before citing them, but they pull from different pipelines: Google layers E-E-A-T checks and Gemini re-ranking on top of its own index, while ChatGPT's browsing mode queries Bing and only shows up in a minority of conversations in the first place. Neither system verifies that a specific code actually works at checkout. For Black Friday shoppers, that means the coupon site an AI answer names is one that passed a trust filter, not one checked against the retailer's live promotions.
What "deciding which coupon sites to trust" actually means
Neither Google nor OpenAI publishes a list of trusted coupon sites. Both systems run a general-purpose source-selection process and apply it to coupon queries the same way they would to a recipe or a product review. Coupons happen to be an unusually spam-exposed category, so the filtering stages that exist for every query matter more here. Understanding how those stages work is the clearest way to predict which page gets cited for "Black Friday [brand] discount code."
How Google AI Overviews select and rank sources
Google's AI Overviews run a multi-stage pipeline that narrows a large candidate pool down to a handful of cited sources. Ziptie's breakdown of the process describes semantic retrieval pulling 200 to 500 candidate documents, followed by an E-E-A-T authority filter that acts as a binary pass/fail gate, then Gemini re-ranking at the passage level, then data fusion into the summary with inline citations. The same analysis, drawing on Wellows' pattern work, found that 96% of AI Overview citations come from sources with strong E-E-A-T signals, while domain authority has become a weaker predictor on its own, with its correlation to citation dropping from r=0.43 to r=0.18.
Candidate pools are broader than the classic top 10 organic results. Ahrefs analysed 4 million AI Overview citations and found only 38% of cited pages ranked in the top 10 for the original query, because Google expands the search into related sub-queries (Paprik's summary of the Ahrefs findings), a mechanism Google itself has confirmed publicly as "query fan-out." Surfer's large-scale study, cited by Bluetree, found around 70% of citations still come from top-10 organic results once fan-out queries are counted, with AI Overview answers showing roughly 8 sources on average.
Passage position inside the page matters too. A CXL analysis of 100 AI Overview citations found 55% of cited passages sat within the first 30% of the page, as reported by Reachbranker's piece on how AI Overviews choose sources, which also notes that content combining text with original images, video or tables was 156% more likely to be selected than text-only pages. Google has also started giving users some direct influence: the company confirmed more than 345,000 unique sources had been selected through its "Preferred Sources" feature across multiple languages and regions, per the same Reachbranker report.
How ChatGPT selects and ranks sources
ChatGPT works differently at the retrieval level. Omnicite's data shows only about 18% of ChatGPT conversations trigger a live web search at all, meaning most answers, including many coupon questions, come straight from training weights with no citations attached. When ChatGPT does browse, it discards most of what it reads: Omnicite's writeup of AirOps research found about 85% of retrieved pages never appear in the final cited answer, which is one reason most of ChatGPT's links don't function as true citations the way a search result does.
ChatGPT's web queries route through the Bing API rather than Google, according to Detekia, then get filtered, reordered and synthesised by the model. Ziptie's separate estimate of browsing-mode weighting puts domain authority at roughly 40%, content quality at roughly 35%, and platform trust at roughly 25%, returning 3 to 6 clickable citations per response. Wikipedia dominates that pool: it accounts for 7.8% of all citations in browsing mode and captures 47.9% of citations among the top 10 sources cited, per Ziptie's analysis. Sutra Creative also reports an 87% correlation between ChatGPT's browsing-mode citations and Bing's own top 10 rankings when browsing is active.
Why coupon and deal pages are a special risk category
Coupon pages carry a documented accuracy problem that predates AI answers. A 2018 study found 90% of 1,051 promo codes marketed on a coupon website over two months were fake, expired or invalid, and the same site promoted codes for 28 affiliate advertisers that don't support coupon codes at all, according to DataDome's research on coupon site bot abuse.
Part of the problem is structural. SimplyCodes' analysis describes a perverse SEO incentive: publishing "No Codes Available" reads to Google as thin content, while publishing "23 Active Codes," even fake ones, reads as a comprehensive resource. Sites that honestly report no working codes can rank worse than ones listing fake ones. The same analysis warns this contaminates AI training data too: if a model has seen "SAVE20" paired with "Nike" thousands of times in scraped SEO spam, it learns that pairing as a statistically likely answer regardless of whether any code ever worked.
Google's 2024 crackdown on third-party coupon content
In early 2024, Google added news sites "hosting coupons provided by a third party" to its definition of spam, while clarifying that publishers actively involved in producing their own coupon content didn't need to worry, as reported by Hello Partner's coverage of the site reputation abuse update. The enforcement hit hard and fast: coupon pages from Daily Mail, Metro and The Mirror, which had dominated coupon search terms, disappeared from Google Search almost overnight.
The reshuffle changed who ranked where. Coupon pages from outlets like the Washington Post, USA Today and the Wall Street Journal had previously outranked dedicated deal sites like RetailMeNot and Groupon, but Google's penalty removed that advantage, as described by Coupons in the News' report on the crackdown's aftermath. Some big publishers, including Forbes, the Wall Street Journal and the Washington Post, deactivated their coupon pages entirely afterwards, worried the penalty would spill over into wider site rankings.
What signals make a coupon page more likely to get cited
A few patterns recur across both engines. Strong E-E-A-T signals matter heavily for Google's filter stage, and content with original images, video or tables is far more likely to be selected than plain text, per the CXL and Reachbranker data above. Passage position near the top of the page correlates with citation in both AI Overviews (55% of passages in the first 30% of the page) and, separately, in a study of over 18,000 ChatGPT citations that found a similar concentration near the top of the document. First-party involvement in the coupon content itself, rather than syndicated third-party feeds, is exactly what Google's 2024 policy rewarded.
It's worth remembering Google treats structured coupon data separately from organic or AI citation entirely. Shipping promotions in Merchant Centre require a valid redemption code and a 6-month validity limit, and the coupon rich result is driven by the Promotions feed rather than on-page schema, per Google's own Merchant Centre documentation and this summary of Magstags' notes on ecommerce rich results. That feed-based path is distinct from anything AI Overviews or ChatGPT cite, so having a Merchant Centre feed doesn't automatically help a page surface in an AI answer.
What this means for shoppers and site owners this Black Friday
For shoppers, the practical takeaway is that an AI-cited coupon site passed a trust filter for authority and content quality, not a live check that the code redeems. Given how AI Overviews already cover a large share of buying-intent searches, as explored in Sited's piece on AI Overviews and buying-intent coverage, shoppers searching "Black Friday [brand] promo code" should expect to see an AI summary before a ranked list of coupon sites, and should still verify the code at checkout.
For site owners, the lesson from Google's 2024 crackdown is that first-party involvement in your own deal content, not scraped or syndicated codes, is what survives spam enforcement. Since most of what ChatGPT retrieves during browsing never makes it into the final answer, as Omnicite's AirOps-sourced data shows, chasing ChatGPT citations directly is unreliable. See Sited's explainer on why most of ChatGPT's links aren't really citations for more on that gap. A file like llms.txt won't fix this either, something covered in Sited's piece on why llms.txt won't make AI systems cite you.
Our view: publishing honest "no codes currently available" copy is the right call for trust, even though SimplyCodes' analysis suggests it can cost short-term SEO visibility. The retailers and deal sites that come out ahead over multiple Black Fridays will be the ones building a reputation for accuracy, not the ones gaming the "23 active codes" pattern that got news publishers penalised in 2024.
One real limitation: none of the research here measures whether the codes an AI engine actually cites are valid at the point of citation. The studies describe source selection and authority signals, not code verification, so a highly-cited coupon page can still contain an expired code. Given that ChatGPT's zero-source answers have also been rising sharply, as covered in Sited's report on ChatGPT's zero-source answer jump, an increasing share of coupon-code answers may carry no citation to check at all.
Frequently asked questions
Does Google verify that a coupon code actually works before citing a page in AI Overviews?
The notes describe Google's selection pipeline as filtering for E-E-A-T signals, semantic relevance and Gemini re-ranking, not code verification. There's no evidence in the research that AI Overviews test whether a cited code redeems successfully.
Why did Google penalise coupon pages from major news sites in 2024?
Google classified news sites "hosting coupons provided by a third party" as spam under its site reputation abuse policy, while exempting publishers actively producing their own coupon content, according to Hello Partner's coverage of the update. The enforcement removed coupon pages from outlets including Daily Mail, Metro and The Mirror almost immediately.
Does ChatGPT check the web before answering a coupon code question?
Only in a minority of cases. Omnicite reports that around 18% of ChatGPT conversations trigger a live web search, so most coupon questions are likely answered from training data with no citation at all, and even when browsing happens, about 85% of retrieved pages never make it into the cited answer.
Why do fake coupon codes keep appearing in search and AI results?
SimplyCodes describes a perverse incentive where listing "no codes available" reads as thin content to Google, while listing many codes, even fake ones, reads as a comprehensive resource and ranks better. That pattern can feed into AI training data too, since a code repeated thousands of times alongside a brand name becomes a statistically likely completion for the model even if every instance was spam.
Sources
- Google AI Overviews Source Selection: Reverse-Engineering How AIO Picks Sources
- How Google AI Overviews Choose Citations
- How Google AI Overviews Choose Sources: 2026 Citation Guide
- How ChatGPT Chooses Sources to Cite
- 5 Types of Coupon Fraud & How to Protect Coupon Sites & Apps
- Site Reputation Abuse & SEO Decline: Affiliates Take a Beating in Google SERPs
- Real Coupon Sites Thrive, After Google's Coupon Crackdown - Coupons in the News
- Google eCommerce Rich Results 2026: Complete Guide
Want more Sited in your Google results?
Add Sited as a preferred source on Google (opens in a new tab)

