7 Tactics for Winning Agentic Search & Shopping, Learned From Building Our Own AI Stack and 200+ Client Cases

When you use a tool or AI Tracker, you only see what that tool was designed to show you. But the moment you play with the raw data behind a single ChatGPT or AI Mode answer, a completely different reality opens up: the patterns your competitors never see, because their tools were never built to surface them. Take AI visibility. What even is it? Every tool hands you a number, almost none of them agree, and almost none tell you the truth. That's what we learned building our own AI stack across 200+ client cases and two years of experiments.

In this session, Stijn Bergmans breaks down 7 tactics from those cases, no theory, just what's actually working right now. What does "AI visibility" really mean when everyone defines it differently, and how do you measure Answer Share properly? Why tracking streaks beats any one-off visibility score in a world where every answer comes out different. How to find the needle in the haystack: the handful of influential citations that actually shape the answer, while everyone else counts noise. And how to SERP-analyse an AI answer itself, so you can see exactly why you're in it or why you're not.

Then it gets weirder on the shopping side. Agents search far wider and deeper than anything you'd ever prompt yourself, fanning one query out into dozens of sub-queries you never wrote. That's where the pure gold is hiding, and where you can give your products a real head start: you'll see how to uncover those hidden sub-queries, how to force your products into AI comparison tables, and how to reverse-engineer from a product spec straight back to the prompts that surface it.
You'll walk out with 7 tactics you can use on Monday, and the blueprint for an AI stack you can build yourself, so you stop guessing from screenshots and start seeing the data your competitors can't, in a world where nobody googles the way they used to.

Transcript

1. 7 Tactics for Winning Agentic Search & Shopping Learned From Building Our Own AI Stack and 200+ Client Cases Stijn Bergmans FolloAgency linkedin.com/in/stijn-bergmans-14996b41
2. Here’s some more OG-stuff we don’t use anymore ×
3. A bit more of the same, retired oldies ×
4. And lastly, 3 things we don’t use anymore ×
5. Stijn Bergmans Head of AI Search - FolloAgency • Dutch guy with 12 years of experience in SEO • 8th time BrightonSEO, 1st time as a speaker • 300+ B2C and B2B clients • Nerd → Loves to be creative with raw data • Not afraid to show you that today
6. Actually, I was in all of these stages of GEO NOV 2022 TO 2023 “Have you seen this ChatGPT thing? Interesting” LATE 2023 TO 2024 “It has a name now: GEO? So this is a real thing.” 2025 2026 “Here you have some budget, go and make me visible.” “Okay, you had my budget. Tell me the impact!” WE ARE HERE ×
7. BIGGEST LESSON I’VE LEARNED: We’re f*cked At least, if we keep looking at SEO the traditional way
8. This was like 18 months ago ×
9. ×
10. The search landscape, 2026 ×
11. There is a blind-spot, called the Dark Search Gap ×
12. Desk research 300+ GA4 properties data beats opinion GA4 MCP: “Did LLM traffic and revenue grow year-over-year? And does it now represent a significant share of the total?”
13. 300+ GA4 properties told me this, YoY: Traffic Revenue / Conversions TOTAL TOTAL +143% +215% LLM sessions · 293K  712K LLM Revenue · € 12M  € 4.1 / 8.2M AVERAGE PER CLIENT AVERAGE PER CLIENT +200% x3 Traffic has tripled on average compared to last year. +308% x4 Revenue has quadrupled on average compared to last year.
14. The slice of the pie is getting bigger Last year Now × 3,6 1% 0,89% × 2,5 × 3,7 0,75% 0,59% 0,56% 0,5% 0,25% 0,24% 0,25% 0,15% 0% LLM Traffic LLM Conversions LLM Revenue ×
15. However, 80% of all properties have an LLM share of less than 1% across traffic and revenue 0  0,5% 0,5  1% 1% and higher Traffic Revenue 277 clients · under 1% 81% (was 96% 184 clients · under 1% 78% (was 97% Now Last year 54% 27% 150 clients 75 clients 84% 231 clients 19% 12% Now 58% 106 clients Last year 22% 21% 40 clients 91% 163 clients Only 20% of properties have an LLM share above 1% ×
16. Across 300+ Properties: Here Are the Peak Numbers Peak LLM Traffic Share 5.16% 0% out of 100% total traffic 100% Peak LLM Revenue Share 4.76% 0% out of 100% total revenue 100%
17. GEO is a brilliant piece of football (soccer, for the Americans in the room)
18. De exponentiële groei van AI Search And the striker gets all the love
19. The customer journey lost its click ×
22. Your Analytics is just the tip of the iceberg 1–2% THE CLICK Everything you see in your Analytics platform WATERLINE = THE CLICK DARK SEARCH/IMPACT 98%+ (AI) INFLUENCE Everything in this phase, from showing up in an LLM answer to a video on social media, a mention on Reddit, or visibility on Google SERPs, doesn't get the credit it actually deserves. ×
23. Meanwhile, our manager or client:
24. So, we as SEOs:
25. We need to fix it before AI Mode becomes the default AI Overviews expand on their own Google Search → AI Overview without SERP’s ×
26. LESSON 1: AI Visibility isn’t the new holy grail
27. I’ve got an answer to my question. So now what? Can I even click through? Do I? Do I Google for the brand name? Do I ask a follow-up question? Do I type in a URL directly? Do I postpone my search journey? ×
28. That's why we built our own toolstack
29. FOLLOW-UP ACTIONS AFTER AN LLM ANSWER Clicks through on the AI answer Straight from the LLM to your site. EACH FOLLOW-UP ACTION LEADS TO SOMETHING MEASURABLE AI Visibility E.G. : pet-insurance prompts: 12%  24% LLM traffic & revenue Keeps chatting E.G. LLM traffic and revenue grow by 150% Compares and decides in the chat — comes back later. Branded volume grows E.G. “acmeˮ and “acme pet insuranceˮ 18% Googles your brand in Search Console Branded search — the direct result of visibility. Direct & homepage visits YOU GET AN ANSWER IN AN LLM E.G. direct homepage traffic rises in step with answer share Types in your homepage directly Knows you now — straight to the homepage. THE FORMULA / SUM Impact on business value One correlation — per prompt, per topic. THIS IS HOW YOU JUSTIFY BUDGET SEO/PPC Growth E.G. branded and topic queries land on /pet-insurance — via SEO & Ads Window-shops products on Google Searches the product — and meets you on channels they already know. Self-reported attribution “How did you hear about us?ˮ — captures what analytics miss. ×
30. ×
31. These metrics let you find AI Impact Track AI visibility month over month next to the metrics it should move. The shape of the lines tells you whether GEO is paying off. Correlation: high AI Visibility LLM traffic & revenue Branded searches SEO & SEA uplift Direct traffic AI Visibility LLM traffic & revenue Branded searches SEO & SEA uplift Direct traffic Conclusion 31% of revenue attributed to AI € 476,000 AI-influenced revenue M1 M2 M3 M4 M5 M6 M7 M8 SEO & SEA uplift and Direct traffic: measured on the homepage and on the pages for the topic you're targeting. Visibility converts: it flows through to revenue, even when analytics gives the LLM zero credit. ×
32. That same toolstack gave me so many more ideas!
33. LESSON 2: Being visible and being recommended isn’t the same
34. This is also possible Track AI visibility month over month next to the metrics it should move. The shape of the lines tells you whether GEO is paying off. Correlation: low AI Visibility LLM traffic & revenue Branded searches SEO & SEA uplift Direct traffic AI Visibility Nothing follows: visibility isn't converting Conclusion Visible, not converting. LLM traffic & revenue Branded searches SEO & SEA uplift Direct traffic M1 M2 M3 M4 M5 M6 M7 M8 Also a signal: you're visible on the wrong prompts or pages, or your brand isn't sticking. Time to adjust. SEO & SEA uplift and Direct traffic: measured on the homepage and on the pages for the topic you're targeting. ×
35. Isn’t an answer just a mini-SERP? Google: ‘have fun sorting through these 50+ elements’ ChatGPT: ‘fewer options, but you still pick’ ×
36. This is wat recommended looks like in an LLM THIS TV IS RECOMMENDED THESE TVs ARE ‘JUST VISIBLE’ ×
37. So I’ve extended our Agent Army to check who’s recommended 01 RAW RESPONSE Google AI Mode 02 RAW · JSON BRAND EXTRACTOR Brand extractor 04 ▸ scan answer × brand entities INSERT INTO brands · 5 rows ▊ "keyword": "what are the best INSERT "item_types": ["ai_overview"], "items": [{ PROMPT DATE BRAND PRODUCT best running shoes… 2026-09-18 ASICS GEL‑Nimbus 28 best running shoes… 2026-09-18 Brooks Glycerin Max 2 Which brands appear in this best running shoes… 2026-09-18 adidas Adizero Adios Pro 4 answer? best running shoes… 2026-09-18 Brooks Hyperion Elite 5 best running shoes… 2026-09-18 Brooks Adrenaline GTS running shoes of this moment?", "datetime": "2026-09-18 11:55", state 1 · what is visible in the answer brands THINKING { "se_type": "ai_mode", BRANDS TABLE "type": "ai_overview", "markdown": "The running shoe landscape features entities passed down · nothing extracted twice standout models… ASICS same table · + recommended: yes | no GEL-Nimbus 28 – renowned for luxury shock absorption… 03 RECOMMENDATION AGENT 05 + RECOMMENDED another stellar option is the Brooks Glycerin Max 2… Recommendation agent brands THINKING state 2 · same table, one column added adidas Adizero Adios Pro 4 – a dominant marathon choice… alternatively, the Brooks UPDATE brands SET recommended ▊ Hyperion Elite 5… Brooks Adrenaline GTS – the gold standard for overpronators…" }] } status 20000 · 13.5 s · $0.004 PROMPT ▸ read entities ← brand extractor Who is actually recommended? UPDATE best running shoes… best running shoes… best running shoes… best running shoes… best running shoes… DATE BRAND PRODUCT RECOMMENDED 2026-09-18 ASICS GEL‑Nimbus 28 YES 2026-09-18 Brooks Glycerin Max 2 NO 2026-09-18 adidas Adizero Adios Pro 4 YES 2026-09-18 Brooks Hyperion Elite 5 NO 2026-09-18 Brooks Adrenaline GTS YES ×
38. Now I know what problem to solve. Do you? BRAND 1 Centraal Beheer 2 THE SHORTLIST THE DECISION LIST Brands that make it into the answer Brands the answer actually picks VISIBILITY RECOMMENDATION % of answers the brand appears in % of those answers it is also recommended in 72% 8% ANWB 61% 38% 3 Interpolis 54% 30% 4 Univé 48% 26% 5 Nationale-Nederlanden 43% 22% 6 OHRA 38% 28% 7 InShared 24% 80% 8 FBTO 11% 33% B A Very visible, but almost never recommended → work on sentiment Recommended when it appears, but almost never visible → work on visibility ×
39. LESSON 3: Streaks made my GEO-life easy
40. Nothing has changed. But a lot has changed. So your visibility went from 38% to… 38% MENTIONED YESTERDAY MENTIONED TODAY Commercial ✓ ✓ 95.6 Commercial ✓ Best car insurance for young drivers? 95.6 Commercial ✓ 4 Where can I find the best price for car insurance? 73.3 Commercial 5 Which insurer is the cheapest for a car? 70.9 Commercial 6 Which car insurer has the best price? 66 Consideration 7 Best car insurance for a lease car? 31.6 Commercial 8 Best car insurance for electric cars? 31.6 Commercial ✓ 9 Best car insurance 2026? 31.6 Commercial ✓ 10 Best car insurance for families? 28.4 Commercial # PROMPT POPULARITY ↓ 1 Which is the cheapest car insurance? 100 2 How do I find the cheapest car insurance? 3 PHASE LOST WON ✓ ✓ LOST ✓ WON ✓ LOST WON ✓ ✓ ×
41. I really hated it when this happened This morning 38% → 30% Visibility −8 out of the answer 01 Kicked Brands are still named. Just not yours answer is no longer influential 02 The No brands are named at all. Nobody is carrying citation dropped 03 Your The page that mentioned you left the competitor took your slot 04 ASame citation, same answer shape, a anymore. search results. Your manager “Why are we down?” winning this prompt now. different name in it. model stopped searching answer format changed 05 The 06 The The answer comes from memory now, not A product carousel, map or video from live sources. dropped the AI Overview 07 Google The prompt no longer triggers an AIO, so there is nothing to be in. replaced the text list. prompt set changed 08 Your Prompts were added or removed, and the denominator moved. A percentage tells you how often. A streak tells you whether it holds, and that is the part you can act on. ×
42. Streaks show you exactly where you lose (or win) # • These answers ALWAYS name a brand. Just NEVER ours. • Or: in these answers, you lost your streak PROMPT INFLUENTIAL STREAK OUR BRAND STREAK PROMPT STATUS 1 Which is the cheapest car insurance? 100% 100% 2 How do I find the cheapest car insurance? 100% 0% NEVER VISIBLE 3 Best car insurance for young drivers? 100% 70% Kicked out after 7 STREAK 4 Where can I find the best price for car insurance? 100% 100% Visible 5 Which insurer is the cheapest for a car? 100% 0% NEVER VISIBLE 6 Which car insurer has the best price? 100% 80% Kicked out after 8 STREAK 7 Best car insurance for a lease car? 100% 0% NEVER VISIBLE 8 Best car insurance for electric cars? 100% 100% Visible 9 Best car insurance 2026? 100% 50% Kicked out after 5 STREAK 10 Best car insurance for families? 100% 0% NEVER VISIBLE Visible ×
43. Streaks also help at prompt-level Prompt: Best car insurance for a lease car? Influential Streak: 100% → Our brand streak: 0% These are the citations that shape the answer (or don’t): PROMPT Best car insurance for a lease car? # RUN DATE CITATION 1 22 Sep comparison-site.nl/lease-car-2026 10 10/10 22 Sep Superstar 2 22 Sep consumer-guide.nl/lease-car-insurance 10 10/10 22 Sep Superstar 3 22 Sep leasing-forum.nl/private-lease-insurance 10 10/10 22 Sep Superstar 4 22 Sep competitor-a.nl/lease-car/insurance 3 8/10 22 Sep Reliable 5 22 Sep competitor-b.nl/private-lease/insurance 5 5/10 22 Sep Rotating 6 22 Sep news-site.nl/lease-car-costs-2026 3 3/10 15 Sep Lucky shot 7 22 Sep youtube.com/watch · lease car insurance 1 1/10 20 Sep Lucky shot 8 22 Sep your-brand.nl/car-insurance 0 0/10 — Never there CITATION STREAK CITED IN 10 CHECKS LAST CITED CITATION STATUS THESE ALWAYS SHAPE THE ANSWER (SUPERSTARS) THESE ARE LUCKY SHOTS ×
44. This is why you lose your visibility (!) • • The LLM isn’t using the citation anymore, where I was on → Falling Star. Now it’s using a new citation, but I’m not on it → Rising Star PROMPT Which is the cheapest car insurance? # RUN DATE LAST CITED NAMES US CITATION STATUS 1 22 Sep comparison-site.nl/cheapest 6 6/10 18 Sep ✓ Falling star 2 22 Sep consumer-guide.nl/cheapest 4 4/10 22 Sep Rising star 3 22 Sep price-check.nl/compare-insurance 10 10/10 22 Sep Superstar 4 22 Sep competitor-a.nl/car-insurance 6 9/10 22 Sep Reliable 5 22 Sep news-site.nl/premiums-2026 1 1/10 15 Sep Lucky shot 6 22 Sep your-brand.nl/car-insurance 0 0/10 — CITATION Named us · ran 6, gone since 19 Sep Names competitors · in since 19 Sep Never named us CITATION STREAK CITED IN 10 CHECKS ✓ Never there Window: last 10 checks. The same read works on 20 or 30. ×
45. Now my Claude is telling me the superstars and rising stars SCHEDULED TASK IN CLAUDE Citation shortlist Runs Weekdays, 0800 Reads FolloEngine streaks via MCP Window Last 10 checks, rolling Posts to Slack · #ai-visibility # ai-visibility Claude APP 800 AM Morning. You can influence 21 answers today by getting on 5 pages: 4 superstars you are not on yet, and 1 rising star that just took a prompt from you. ★ Superstars · 19 answers · cited in every check, we are never named comparison-site.nl/lease-car-2026 7 answers · 4 prompts consumer-guide.nl/lease-car-insurance 6 answers · 4 prompts 3 leasing-forum.nl/private-lease-insurance 4 answers · 3 prompts 4 price-check.nl/compare-insurance 2 answers · 1 prompt 1 THE PROMPT Look at the last 10 checks across all our prompts. Give me a shortlist of the citations I need to focus on: superstars that carry answers where we are never named, and rising stars that just took over a prompt we lost. One line per citation, with the prompts it decides. Set once. Claude reads streaks, never raw answers. 2 ★ Rising stars · 2 answers · took over a prompt we held for 6 checks 5 consumer-guide.nl/cheapest 2 answers · in since 19 Sep · replaced comparison-site.nl, which named us Five pages, 21 answers. Everything else can wait. Reply with a prompt and I will show its citations. 👀3 ✅2 ×
46. Streaks are addictive. You want more. STREAK WHAT IT COUNTS WHAT YOU DO Websearch streak The model searched the web for this prompt Win it with pages. Streak breaks: build brand instead. Training data streak The answer came from model memory, no live search No page fixes this. Be present where models train. AIO streak Google shows an AI Overview on this prompt AIO stable: write for the overview. AIO gone: chase rankings. Influential streak The answer names or ranks brands Spend budget where brands get recommended. Our mention streak Your brand is named in the answer Answer “why are we downˮ before it is asked. Recommended streak Your brand is the one the answer recommends Mentions get you listed. Recommendations get the customer. Our citation streak Your own domain is cited At 0, your mention depends on someone else. Citation streak One specific source in the last web searches for prompt Influence the top three pages. Ignore the rest. Product carousel streak The answer shows a product carousel It is a feed, not a text. Fix your product feed. Video streak The answer embeds video Publish the video before someone else does. Community citation streak Reddit or forum pages are cited Show up in the thread, not on a new landing page. Freshness streak Cited pages are younger than 30 days Update inside that window or lose the slot. First-mention streak Your brand is named first Be first. The rest of the list gets skimmed. ≈ ×
47. LESSON/LEARNING 4: Chunks, chunks, chunks
48. Hover on a citation in AI Overviews / AI Mode ×
49. How do you approach you visibility? 12% 14% Almost everyone: I’m cited, so job done 12% 14% 36% Almost no one: I own just 14% the answer while my competitors is owning 36% of the answer! 12% ×
50. Two flows, collecting the data you need A AI ANSWERS → AGENTS CHUNK AND CLASSIFY THEM Chunking agent ▸ split answer → chunks text + references per chunk ▌ B OWN WEBSITE → ALL PAGE CONTENT IS STORED IN WAREHOUSE URLs AI Overviews and AI Mode answers Calculator ▸ chunk chars ÷ visible chars real estate per source ▌ ONPAGE API /advies/puppy-in-huis /advies/brokken-per-dag /kattenvoer/kitten ▸ crawl every URL ▸ extract headings + content Data warehouse Citation classifier ▸ classify → first / third party brand-owned or not ▌ 1,842 website_chunks · 9,318 rows URL ELEMENT CONTENT /puppy-in-huis H1 Puppy at home: the first weeks /puppy-in-huis H2 The first night /puppy-in-huis P Your puppy sleeps best close to you at first… /puppy-in-huis H2 House training /puppy-in-huis LI Go outside after every meal /brokken-per-dag TABLE Grams per day by body weight ×
51. Now you know your real visibility This is what you are probably looking at This is the space you own in all AI Overviews This is the space you own in all AI Overviews you can influence with your own website Answer Share AIO Real Estate First Party AIO Real Estate First party 54% 9% 22% royalcanin.nl First party 37% 18% 25% reddit.com Third party Forum 21% 15% — eukanuba.nl First party 18% 9% 16% hills.nl Third party Corporate) 41% 14% — youtube.com Third party Video 22% 11% — zooplus.nl Third party Retailer) 29% 9% — Source Source Type prins.nl SHARE OF VISIBLE TEXT 55% FIRST PARTY 55% first party brand sites, ours or a competitorʼs 45% third party forums, retail, video, institutions ×
52. And now you can create Real Estate in AIOs 01 SUBJECT STREAK AGENT 02 Subject agent WE HAD OUR OWN CONTENT 03 Coverage agent ▸ cluster → subject · streak · real estate all answers · 50 measurements ▌ Not covered → Create Covered, not shown → GEO-proof ▸ match × website_chunks 5 subjects · 9,318 chunks ▌ · anxiety & stress → streak high · 9.4% real estate JUST A FORMULA Covered and shown → Keep · weight → covered · not shown Subject Streak Real estate Covered? Shown? Action Anxiety & stress High 9.4% No No Create Weight & portions High 6.1% Yes No GEO-proof Dental care Medium 5.8% No No Create Switching food Medium 4.2% Yes Yes Keep Grain-free food Low 2.0% Yes No GEO-proof Streak: how consistently the subject appears across measurements. Real estate: share of all visible text it takes. Priority = high streak and high real estate. Covered: the subject exists in our own content. Shown: our content is actually cited for it. GEO-proof = rewrite existing content so an LLM can cite it: heading, list, direct answer. ×
53. LESSON/LEARNING 5: Win Google Shopping to win AI Shopping (yes, also ChatGPT, for now)
54. ChatGPT’s response betrays (betrayed) itself It searches organic Google Shopping, until it has an own product database of it’s own ], "product_ids": [ { "type": "chat_gpt_google_shopping_product", "ei": "nw0parmtO-0ww8cPkJCG4Q0", "product_id": "3385313524518459355", "catalog_id": "5258429276028622309", "gpcid": "14086073946244683030", "headline_offer_docid": "5439354036607622721", "image_docid": "11189928386663626884", "query": "beste+oled+televisie+lg+g5+samsung+s95f+sony+bravia+8+ii", "pvt": "hg", "uule": "w+CAIQICIfRW5zY2hlZGUsT3Zlcmlqc3NlbCxOZXRoZXJsYW5kcw", "gl": "nl", "hl": "nl" } ] } ×
55. LESSON/LEARNING 6: Raw data is a goldmine
56. ChatGPT and Google tell you what the model compares on ×
57. Query fan-out: what the model adds that you never typed Best mattress for back pain ? I sleep hot , my Best mattress for back pain ? I sleep hot , my partner tosses and turns, budget around €800 ? partner tosses and turns, budget around €800 ? Query Fanout Query Fanout mattress back pain cooling gel pressure relief sleep hot best mattress hybrid back pain sleep hot Shopping Query Fanout Shopping Query Fanout best mattress mattress €800 medium-firm motion isolation orthopedic €800 firm pocket spring zoned support ×
58. The best part: you can extract them ChatGPT · table DataForSEO LLM Scraper "table": { "table_header": [ "Shoe", "Best suited to", "Cushioning", "Speed", "Stability", "Feel", "Approx. UK price*" ], "table_content": [ [ "**ASICS Novablast 5**", "Everyday + long runs", "\u2b50\u2b50\u2b50\u2b50\u2b50", "\u2b50\u2b50\u2b50\u2b50", "\u2b50\u2b50\u2b50\u2b50", "Bouncy/energetic", "~\u00a390\u2013\u00a3160" ], ... ] ChatGPT · fan-out // request [{ "keyword": "best running shoes for a marathon", "location_code": 2826, "language_code": "en", "force_web_search": true }] // response "fan_out_queries": [ "best running shoes marathon 2026" ] DataForSEO LLM Scraper AI Mode / AI Overviews · table DataForSEO LLM Scraper "table": { "table_header": [ "Shoe Model", "Category / Type", "Cushion Level", "Heel-to-Toe Drop", "Best For", "Key Tech / Foam" ], "table_content": [ [ "Nike Pegasus 41Go to product viewer dialogue for this item.", "Neutral Daily Trainer", "Medium / High", "Moderate (10 mm)", "Everyday road runs, versatile training", "ReactX foam + Air Zoom units" ], ... ] AI Mode / Overviews · fan-out Gemini 3.5 Flash · web search // request { "model": "gemini-3.5-flash", "input": "best running shoes for a marathon", "tools": [{ "type": "google_search" }] } // response step (illustrative queries) { "type": "google_search_call", "arguments": { "queries": [ "best marathon running shoes 2026", "carbon plate marathon shoes review" ] } } ×
59. Use these ninja subjects on your PDP and in your Feed(s) ×
60. LESSON/LEARNING 7: Golden Record Feeds really do work
61. A · 60% FEED 60% B · GOLDEN RECORD 99.9% × product-1 product-2 product-3 product-1 product-2 product-3 Required gets you listed. Optional gets you recommended. Most feeds we see fill the required 15, plus a color here and there: a 60% feed. The optional attributes, the ones AI reads to understand a product, stay almost empty. custom_label_1 custom_label_0 return_policy_label energy_efficiency_class certification loyalty_program unit_pricing_measure sale_price shipping_weight product_weight product_height product_width product_length virtual_model_link lifestyle_image_link additional_image_link pattern color material google_product_category Required · 15 product_type short_title popularity_rank variant_option item_group_title document_link related_product question_and_answer product_detail product_highlight shipping size item_group_id identifier_exists condition mpn gtin brand availability price image_link link description title id Most feeds stop at what Google requires Optional · 30 and counting
62. Enrich with the right data, not just more data GOOGLE SHOPPING FEED today, and what ChatGPT Shopping reads as well Purple fanout terms Lilac AI comparison table columns Attribute A · 60% feed B · Golden Record title Mattress 140200 Pocket Spring Cold Foam 7 Zone H3 Mattress Hybrid mattress 140200, medium-firm with cooling gel, for back pain description Pocket spring mattress with 7 zones. Height 24 cm. Firmness H3. Free delivery. Best suited to back and side sleepers with back pain. Firmness: medium-firm. Support: zoned support from pocket springs for pressure relief. Cooling: a cooling gel top layer if you sleep hot. Motion isolation, so a restless partner doesnʼt wake you. product_highlight empty Zoned support that eases back pain Cooling gel layer for people who sleep hot Motion isolation for restless partners product_detail empty Comfort · Best suited to · Back and side sleepers Comfort · Firmness · Medium-firm 6/10 Construction · Support · 7-zone pocket spring, orthopedic Comfort · Cooling · Cooling gel top layer question_and_answer empty Q Is it good for back pain? A Yes: medium-firm with zoned support for pressure relief. Q Does it sleep cool? A Yes, a cooling gel top layer for people who sleep hot. Q Can I try it first? A Yes, a 100-night trial at home. new · conversational SOON CHATGPT PRODUCT FEED ChatGPT moves to its own feed once it has enough product data to answer reliably. With different metrics than Google Shopping: popularity_score q_and_a return_rate raw_review_data product_review_rating enable_checkout ×
63. The A/B test: Golden Record feeds win TEST SETUP VARIANT A · CONTROL Standard feed ± 60% attribute completion, keyword titles, batch updates ChatGPT Shopping Google Shopping Recommendations per 1,000 shopping prompts Impressions in Shopping & AI Overviews +40% +50% more recommendations with Golden Record VARIANT B · TEST Golden Record 99.9% complete: intent copy, highlights, details, real-time price Same products, same prices. 2,140 recommendations more impressions with Golden Record 2,996 276K recommendations impressions level A 184K level A impressions Catalog split 50/50 · 6 weeks OUTCOME +32% +27% more clicks more revenue from Shopping and AI answers on the Golden Record half A · 60% FEED B · GOLDEN RECORD A · 60% FEED B · GOLDEN RECORD ×
64. Wrap up: 7 things to do Monday morning 01 Visibility isn't the KPI. 02 Split shortlist from decision list. 03 Report streaks, not percentages. 04 Being cited isn't owning the answer. Track the follow-up actions next to it: LLM traffic, branded search, direct and SEO/PPC uplift. Visible but not recommended is a sentiment problem, not a visibility problem. Answer “why are we down?ˮ before it is asked. Chunk the answers and measure the real estate you own. 05 Win Google Shopping first. 06 Read the raw response. 07 Fill the optional feed attributes. ChatGPT Shopping runs on it. For now. Fan-outs and table columns are the model's comparison criteria. Required gets you listed. Optional gets you recommended. Your analytics shows you 1 to 5%. Go measure the rest. ×
65. Thanks! Any Questions? Ask ChatGPT, or ask me 😉