Scores Every Webpage Inside the web retrieval pipeline of GPT-5.6 and GPT-6 "@retrieval_score": 113.88 "@retrieval_feature:recency": 100 "sonic-re-fh-chunk-ev3-sx4": 0.9999 "minhash_dedup": 0.38 Metehan Yeşilyurt · Peec AI
AI I am a GEO Researcher at Peec AI and a digital marketing professional with over a decade of experience in the industry. Drop your photo here I work on Generative Engine Optimization (GEO), Answer Engine Optimization (AEO), and SEO. I research how Large Language Models and AI search engines find, rank, and cite content. I am a recurring speaker at major global search conferences and a contributor to publications such as Search Engine Land. I am also a proud father of a daughter. SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026
tobacco towns The Black Sea coast 9,155 km BLACK 19 Mayıs, Samsun SEA where I grew up Durham, NC TechSEOConnect Istanbul Tobacco country 19 Mayıs is a tobacco district. Tobacco Europe's 2nd largest tobacco factory Durham, North Carolina : A city built on tobacco Bull Dukes, the American Tobacco has shaped my life there for British American Tobacco runs it today, the generations. in my home district. Company. Metehan Yeşilyurt · Peec AI · TechSEOConnect · Durham, NC Durham, 19 Mayıs Trabzon Why this stage matters From one tobacco town to another. Being here means a lot to me.
score We read two full web search threads, one on GPT-5.6 and one on GPT-6, field by field. 01 The setup 02 Query fan-out 03 The scoring formula 04 From results to prompt 05 Citations and tests 06 What to do
inside the SSE stream. 01 Found Sep 2026 Every web search answer carried a hidden debug payload: queries, scores, prompt blocks, A/B tests. 02 Studied Sep to Oct 2026 Hundreds of SSE dumps, collected from the US, Europe and Asia. 03 US Europe Asia Each dot is one SSE dump. The blue one is the lens. SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026 About one month later OpenAI closed the leak. What follows is a record of that window. Why the numbers kept moving ChatGPT runs a very large number of A/B tests. Our findings shifted from dump to dump, so we read hundreds of them. Patched Real numbers, one lens 1 SSE ANN, BM25, recency, RRF and every other value on these slides is real, taken from a single SSE dump per model. The patterns come from the hundreds.
Sep 25, 2026 · 3 turns Oct 12, 2026 · 8 turns What I typed What I typed “Give me the most exhaustive query level research map “what are the best ai visibility tools” of the best GEO tools market.” "model_slug": "build_sha": "gpt-5.6-luna", "model_slug": "cdd88ff2", "eval_preset": "build_sha": "sonicberry-unpaid-oai-recency", "completion_model": "gpt4o-sonic-1p-ev3" "gpt-6-luna-api-ev3", "6049ee64", "eval_preset": "sonicberry-unpaid-oai-recency", "completion_model": "gpt4o-sonic-1p-ev3" 23 322 56 25 331 39 search queries results scored domains, turn 0 search queries results scored domains, turn 0 SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026
10 1 search engines configured switched on for this user fired on every query index behind them all "search_engines": {...65} "enabled": true "atc_query_debug_info" "source": "labrador" All 10 engines were Labrador variants: web, reddit, pdf, wiki, arxiv, youtube and four news windows. Bing, SerpApi and Oxylabs sit in the config but did not fire on these turns. SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026
84% Side tools in parallel: products, local, availability 7% of the answer timeline is retrieval: fan-out, scoring, reranking. Safety clear and fetch pages 1% Prompt render and main model 13% GPT-5.6 turn: 83% GPT-6 turn: 84% Citations and logging 1% 0% 25% 50% GPT-6, turn 0. Steps from search.dag.step.duration, shown as share of the full timeline. SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026 75% 100%
query The model rewrites every question before the index sees it. 01 The setup 02 Query fan-out 03 The scoring formula 04 From results to prompt 05 Citations and tests 06 What to do
visibility tracking tools 2026 Competitor A Competitor B ai visibility toolkit Competitor C ai visibility What I typed “what are the best ai visibility tools” +8 query 2 system2_search_query ai visibility monitoring tools 2026 best citation tracking chatgpt gemini perplexity prices recency: 30 query 3 brand names added that I never typed + the year “2026” in 3 of 3 queries SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026 best ai visibility monitoring tools 2026 peec ai Competitor D Competitor E Competitor F Competitor G ai search tracking comparison
Queries in turn 0 4 3 Quoted “exact phrases” in turn 0 4 of 4 0 of 3 Recency window in turn 0 3650 days 30 days The year “2026” in the query 1 of 23 14 of 25 site: queries on brand domains, all turns 15 of 23 20 of 25 Average words per query 7.2 7.4 Distinct brands named in queries 12 11 Highlighted rows changed between models. Same retrieval stack, different query writer. SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026
number And the number is simpler than you think. 01 The setup 02 Query fan-out 03 The scoring formula 04 From results to prompt 05 Citations and tests 06 What to do
technologyadvice.com Scoring What each score means, in plain words @rrf { "sonic_search_engine": "system2-web", "score": 6.9999, "score_debug": [ ["@rrf", 6.9999], ["sonic-re-fh-chunk-ev3-sx4", 0.99998], ["minhash_dedup", 1], ["sonic-emb-cotrain-ev3", 0.7396], ["@retrieval_score", 113.88], ["@retrieval_feature:bm25", 1.44], ["@retrieval_feature:ann", 12.41], ["@retrieval_feature:log_in_search_result", 0.03], ["@retrieval_feature:required_phrases", 0], ["@retrieval_feature:domains", 0], ["@retrieval_feature:recency", 100] ] } Final fused rank across all queries and engines. sonic-re-fh-chunk-ev3-sx4 Reranker score for each text chunk of the page. sonic-emb-cotrain-ev3 Embedding match between query and page. minhash_dedup 1 means unique. Lower means near duplicate. @retrieval_score The plain sum of the features below it. ann / bm25 Meaning match / keyword match. recency 100 if the page sits inside the time window. SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026
is a plain sum recency 100 required_phrases +0 ann bm25 log_in_search_result + 12.41 + 1.44 + 0.03 @retrieval_score = 113.88 Share of the score recency 88% GPT-6, turn 0, the first web result. Same sum on 62 of 62 results in GPT-6 and 77 of 77 in GPT-5.6. SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026 ann 11% bm25 1%
ann Scoring bm25 Who gets the 100 120 113.9 90 Only web Inside one engine, a fresh web page starts 100 points ahead. Reddit, PDF, wiki and arXiv got 0 recency points in both dumps. 60 30 12.2 web web-pdf 0.7 0.5 0.5 reddit arxiv wiki SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026 0 "system2-web": { "rewriter_use_recency": true, "recency": 30 }
no quotes required_phrases ai visibility monitoring tools 2026 best citation tracking... +100 for each quoted phrase in the query that the page matches. GPT-5.6 wrote quotes in every turn 0 query. GPT-6 dropped them. 113.8 100 GPT-5.6 · 1 quoted phrase "ai search visibility" tools geo aeo tracking platforms 100 +100 213.0 GPT-5.6 · 2 quoted phrases "generative engine optimization" tools "ai visibility" platform tracking 100 +200 Same kind of page. Same meaning score. Up to 2.75x the retrieval score. SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026 313.1
turn 0 · 36 web results in the prompt 24 of 36 left out 2.10 made the prompt. ann ranges 11.6 to 13.0. bm25 ranges 1.3 to 2.0. There is no score threshold. bm25 (keyword) 1.88 1.65 3 of the 12 left out were duplicates of pages already in. The other 9 ranked lower after fusion. 1.43 1.20 11.5 11.9 12.3 ann (meaning match) SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026 12.7 13.1
reddit arxiv Fusion wiki 7.10 after rank 36: nothing reached the prompt 6.52 5.95 5.38 4.80 all 24 prompt slots came from this flat zone rank 1 "rrf_halflife_position": 12 rank 62 "rrf_override_result_position": 8 SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026 "score_scale": "normalized"
Chunks Snippet score Prompt your URL hybrid parser # h, lists, tables ~178 words each one per chunk wordlim 200 GPT-6 · turn 0 · technologyadvice.com · 2 chunks # Best AI search monitoring tools for 2026 Compare the best AI search monitoring tools for 2026, including Competitor D, Competitor F, Competitor E ... 1.0 160 words 1,020 chars Written By ... Published: Jun 23, 2026 Updated: Jun 24, 2026 Some tools focus on prompt-level visibility and citations, while others add competitive benchmarking ... ## Best AI search monitoring tools comparison Provider | Best for | Key features --- | --- | --- | --- chunk 0 | My score SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026 chunk 1 0.9994 253 words 1,790 chars 1,074 178 397 to 2,116 max 366 64% 34 of 153 210 of 326 turn 0, GPT-6 chars in a median chunk of pages: first chunk scored 1.0 words in a median chunk later chunks started at a heading Parser: html2text, in a hybrid setup We have seen OpenAI update and improve this layer again and again. It will change. Clean HTML that turns into clean Markdown survives every version.
sonic-re-fh-chunk-ev3-sx4 · 214 chunk scores Rerank Chunks per fetched page 1 61 68 19 2 7 3 56 16 4 5 6 15 < 0.90 14 0.90 to 0.99 0.99 to 0.999 0.999 to 0.9999 ≥ 0.9999 Median chunk score: 0.9992. The reranker is almost always sure. The gap that matters is in the last decimals. SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026 6 7 3 10 3 11 1 15 1 61 pages became 214 chunks. Each chunk is a [text, score] pair in page_contents.
on all four: “Competitor A: See ANY brand’s AI visibility” Index copies competitor1.com/competitor-page?gad=1 minhash_dedup 0.74 in prompt minhash_dedup 0.46 in prompt competitor1.com/competitor-page?tp=1 competitor1.com/competitor-page?authuser=0 minhash_dedup 0.46 competitor1.com/competitor-page?from=groupmessage&isappinstalled=0 minhash_dedup 0.38 dropped dropped Same story on competitor10.com: /en and /en-us versions, dedup 0.52, both in the prompt. SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026 4× Tracking and app parameters created four documents for one page. Two still took prompt slots.
to 24 prompt blocks Retrieval finds you. The prompt budget decides if the model reads you. 01 The setup 02 Query fan-out 03 The scoring formula 04 From results to prompt 05 Citations and tests 06 What to do
web reddit pdf news GPT-5.6 · retrieved 48 GPT-5.6 · in prompt 22 GPT-6 · retrieved 36 GPT-6 · in prompt 24 arxiv Prompt wiki 12 6 4 100% 100% 12 Every prompt slot in turn 0 went to a system2-web result. Reddit, PDF, wiki, arXiv and news were fetched and scored, then left out. SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026 6 6 4 4 100% 100%
search result block Prompt 1 The citation handle. turn, type, index. <|im_start|>search_results to=assistant sonic_webpage<|im_sep|> Best AI search monitoring tools for 2026 | TechnologyAdvice (https://technologyadvice.com/blog/.../best-ai-search-monitoring-tools/) 【turn0search0】 [wordlim: 200] Published: 3 months ago; Crawled: 2 weeks ago; 2 [wordlim: 200] Hard cap on words shown from your page. Available data binding: turn0search0: WebRef[title, timestamp, domain, url] # Best AI search monitoring tools for 2026 Compare the best AI search monitoring tools for 2026, including Competitor D, Competitor F, Competitor E, Competitor H, Competitor I, Competitor A, Competitor G, and Competitor B ... <|im_end|> 【turn0search0】 3 Published / Crawled Your date. If there is none, the crawl date. 4 WebRef data binding New in GPT-6. Title, time, domain and URL as a typed object. SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026
page GPT-6 GPT-5.6 of the page reached the prompt of the page reached the prompt Page Page In the prompt In the prompt 5.8% 1,523 tok 88 tok Each square is 1% of an average page. num_tokens on the result vs ntokens on the prompt block (wordlim 200). SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026 4.6% 1,528 tok 70 tok
labels in the 24 GPT-6 prompt blocks Blocks with a publish date Published: 3 months ago Published: last month 3 Published: last week The other 11 showed only when Labrador last crawled them. No date on the page means the crawl date is your age. 2 Published: 2 months ago 1 Published: 4 months ago 1 Crawled: last week 7 Crawled: today 6 Crawled: 2 weeks ago Crawled: yesterday 13 of 24 6 5 3 SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026 System line above every result "Today is Monday, October 12, 2026 at 12:55:53 PM in the time zone Europe/Istanbul. The user is in Ankara, Ankara, Türkiye."
and who gets tested Attribution settings and the A/B tests running on this account. 01 The setup 02 Query fan-out 03 The scoring formula 04 From results to prompt 05 Citations and tests 06 What to do
Attribution settings, same in both builds 50 Domains for reasoning samples 80% A ranked domain list is built for the model’s chain of thought. GPT-6 turn 0 baseline 0.02 threshold 0.45 p = 1 ranked 39 domains, with a top 80 percent setting. enable_attribution attribution_n_samples true 50 attribution_prob_threshold attribution_prob_baseline 0.45 0.02 "cot_search_result_ranking": "cot_search_result_domains _top_percent": attribution_influence_temperature attribution_sampling_method 0.2 "koo" SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026 80 [39],
Model GPT-5.6 + GPT-6 Build diff cdd88ff2 → 6049ee64 emit_model_data_binding_declarations shopping_nr_opd_postrank_v44_20260922 "" GPT-5.6 false → true WebRef binding filter_serp_by_robots shopping_sid_hybrid_h200_r50_v1 "" GPT-5.6 false → true robots.txt checked budget_trimmed shopping-better-offer-matching Test GPT-6 new field on each result allow_single_image_groups Layers: sonic_chatgpt_paid_ab_tests · sonic_search_engine_all · product_search_all → true images labrador-images base_index_only → true eval_preset, same in both image retrieval_model sonicberry-unpaid-oai-recency imgemb_v2 SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · S ep to Oct 2026 images images
to the code 01 02 03 Show a real date Front-load the first 200 words Be a web page, not a thread In window gets +100 recency. Undated pages show only “Crawled”. The model sees about 88 tokens of a 1,500 token page. Reddit got 24 retrievals and 0 prompt slots in GPT-6 turn 0. recency +100 wordlim 200 system2-web 04 05 06 Keep one URL per page Make your site worth a site: query Use the exact words people quote 20 of 25 GPT-6 follow-ups targeted a single domain. Each quoted phrase that matches adds +100 in GPT-5.6. Variants get merged by minhash. Canonicals and clean params win. minhash_dedup site: 20/25 SSE debug dump · ChatGPT web search · GPT-5.6 and GPT-6 · Sep to Oct 2026 required_phrases
visibility tools WINS LOSES brand-blog.com/resources/guide?utm_source=newsletter&ref=nav example.com/ai-visibility-tools Best AI visibility tools in 2026: 9 tools compared The Ultimate Guide to Navigating the AI Landscape What html2text hands to the chunker Accept # The In cookies Ultimate today's does [image: to moving mean of for large | Solutions Navigating digital intelligence that rise Products Guide fast Artificial the | your What html2text hands to the chunker is Resources the AI world, changing business? language | models In and face people this why Contact sales # Landscape brands how | new we AI visibility Published: challenges search, guide Best shop explore and visibility every the matters day. discover. history more than of But search, 2, 6 in 202 6 · Updated: Oct 9, 2026 202 6 are Brand A for prompt tracking, Brand B for citation analysis and Brand C for agencies. We tested 9 tools on 500 prompts in ChatGPT, Gemini and Perplexity. The what Oct tools best AI visibility tools in 202 ever... comparison-table.png] | Tool | Best for | | --- | --- | --- | Price | from $89/mo | | Brand B | Citation analysis | $129/mo | | What the model reads 【turn0search17】 Accept # The Ultimate | Products today's fast no date 4 Guide In to moving URL A | Prompt tracking | What the model reads [wordlim: cookies Brand 200] | Crawled: Solutions Navigating digital variants | the world, table 2 weeks Resources AI an | 【turn0search0】 Contact sales # Landscape brands is ago; face image Best The new challenges every day... wer after word 600 ans A made up example · built from the rules in this deck · not a real page AI best citation dated visibility AI analysis 200] tools visibility his week t [wordlim: and 6 in last ; Crawled: today; week 202 tools Brand one Published: in C 6 are Brand A for prompt tracking, Brand B for 202 for agencies... canonical URL real HTML table wer in sentence 1 ans
Losing page Winning page Recency No date on the page. Risks the window and shows only “Crawled”. Published last week. Inside the window: +100. recency +100 Phrase match Says “AI landscape”. Never says “ai visibility”. “AI visibility tools” in the H1 and sentence 1. required_phrases Dedup Same page under 4 tracking URLs. Signals split. One canonical URL, clean parameters. minhash_dedup Parse and chunk Chunk 0 is a cookie banner and nav. Table is an image. Chunk 0 is the H1, the answer and a Markdown table. html2text irst 200 words A history of search. The answer starts after word 600. Names, use cases and prices in the first 80 words. wo ollow-up Pricing says “Contact sales”. The site: query finds nothing. / site: 20/25 itation Read, then left out of the answer. Clears the attribution test. Cited. t F F C Retrieved, scored, maybe read. Not cited. Signal pricing lists every plan with a number. rdlim 200 hreshold 0.45 【turn search 】 in the prompt ited in the answer 0 0 .C .
AI Peec AI · TechSEOConnect · 2026 Email [email protected] Newsletter metehanai.substack.com X x.com/metehan777 LinkedIn linkedin.com/in/metehanyesilyurt Site metehan.ai
metehanyesilyurt
charlesmeaden
amanaut
speedyindexteam
joshbly
.jpeg){kind=avatar}
pelogia
samtorres
dawnieando
gkoivikko
.png){kind=avatar}
samanyougarg
katarinadahlin
aleyda
soysaucechin
lcolladotor
speakerdeck
codingconduct
lynnandtonic
caitiem20
davidcarrasco
coloredviolet
sugarenia
sarafernandez
jonoalderson
inesmontani
