← Back to selected work

JIA JINGQIU / METHODS & TOOLS

GEO, in practice.

Google-account pools, channel-specific publishing, and clean ChatGPT / Gemini recommendation observations.

04 / GEO & DISTRIBUTION

GEO: distribution
and recommendation analysis.

My GEO approach has two connected tracks: Google-account pools and Agents organize channel-specific publishing; repeated ChatGPT and Gemini observations show which products, sources and conditions appear. The next change is tied to a concrete content gap or channel task.

WALK THROUGH A DECISION / DEMO

A monitor arm was not recommended.
What should happen next?

DEMO-ARM, its D27 product and the answers below are teaching examples—not actual ChatGPT logs or store results. Specifications, prices and supply still need real evidence. “27-inch” in a name does not prove compatibility.

Fix the question. Keep the failures.

In the US, which desk arms under $150 suit two 27-inch monitors? Explain compatibility requirements.

SAMPLE SCOPE
US / English / question v1 / one consumer-product surface and matched session conditions, with six planned runs. In real observations, record the interface, visible model, personalization and search state per run. Keep ChatGPT, Gemini and API results separate.
ILLUSTRATIVE ANSWER R01
“Compare two alternative arms. Check weight, mounting interface and desk conditions, not just screen size. DEMO-ARM D27's fit still needs verification.” This illustrative answer cites target compatibility page v1 but does not recommend D27. The page is a simulated source, not a claimed existing URL.
WHAT IS OBSERVED
This answer does not recommend the target product; its visible citation concerns compatibility. Preserve the answer, citation location and page version, then inspect the product evidence.
WHAT REMAINS UNKNOWN
This does not establish that the model cannot recognize the brand, that one channel has a higher weight, or that more posts would secure inclusion. Visible explanations are not internal ranking logs.
DEMO before: six planned runs and four valid answers
RunStatusTarget recommendation / citationSearch confirmed
R01Valid; target product mentionedNo / YesYes
R02Valid; target product mentionedYes / NoYes
R03Valid; target product mentionedYes / NoNot confirmed
R04Valid; target product not mentionedNo / NoNot confirmed
R05Technical failure; no valid answerOutside the valid-answer denominatorNot assessed
R06Planned; no capture recordRetained as unobservedNot assessed

Merge duplicate imports of the same answer, but retain genuinely separate runs even if their text matches. This group has 2/4 recommendations and 1/4 target citations. The confirmed-search subset has two valid answers and one citation: report 1/2 separately. A recommendation does not establish that specifications or prices are correct.

Mechanisms and measurement scope used in this example

Supporting links in Google AI Overviews / AI Mode require indexing and snippet eligibility. This applies to Google's Search features, not every answer product. Google AI Search guidance ↗

ChatGPT shopping treats product selection and merchant offers separately, and price or shipping updates may lag. Check referrals in actual analytics, including utm_source=chatgpt.com where present; do not extrapolate site clicks from sampled answers. Shopping results ↗ Publisher and referral guidance ↗

Bing AI Performance's citations, URLs and queries cover supported surfaces, with some data sampled. They are not total web-wide answer volume or channel weights. Bing measurement scope ↗

01 / ACCOUNT POOL & AGENT PUBLISHING

More accounts need clearer roles.

Register Google identities, platform publishing permissions and browser/mobile environments separately. Assign brand, market, language and content purpose, then give the Agent a specific task.

01Google accounts & permissions

Record role ID, represented brand/team, owner, market, language, available services and status. Google sign-in is not permission to publish everywhere; use existing authorized accounts.

02Browser & mobile

Bind browser tasks to the correct AdsPower environment and native-app tasks to an authorized device. Verify the actual identity and destination; keep credentials and device/proxy details private.

03Queue & submission

Each task has a topic, channel, destination, content version, approved scope and owner. Prefer existing APIs/connectors; use permitted browser/device flows when a suitable API is unavailable.

04Readback & exceptions

Avoid concurrent writes to one destination. Reopen and check text, media, links and visibility. Inspect uncertain submissions before retrying; hold permission failures, challenges and rate limits.

Account count is execution capacity, not independent users or a ranking weight. Keep publishing accounts separate from measurement sessions. Do not rotate identities to evade restrictions or manufacture customer reviews.

PUBLISHProduct factsChannel / AccountPlatform versionApproved queueAPI / Browser / MobilePublic readback

02 / CHANNEL MAP

Name the channel. Define its job.

Combine observed citations, owned-content gaps, publishing conditions and production cost when assigning tasks. A cited source is not necessarily open to submissions. This is a candidate list, not an algorithmic weight table or a requirement to use every channel.

01 / Owned material

Brand website · Product / comparison / FAQ pages · Blog · Help / developer docs

Work to doCompletion boundary
Specifications, compatibility, prices, limitations, original demonstrations and instructions. Maintain one factual source for channel versions.Authorized owned publishing; access, indexing and citations need separate checks.
02 / Google properties

YouTube · Blogger · Google Sites · Google Business Profile

Work to doCompletion boundary
Demonstrations with text, focused pages, brand facts and local business information.A Google account is account context, not universal publishing permission. Business Profiles require an eligible real business.
03 / Public articles

LinkedIn Articles / Newsletter · Medium · Substack

Work to doCompletion boundary
Use-case tutorials, comparisons and buying decisions with facts, examples and limitations.Separate public articles from paywall previews and email; verify identity and editor permissions.
04 / Communities & Q&A

Reddit · Quora · Hacker News / Show HN · Indie Hackers · Stack Overflow · Stack Exchange · Industry Discourse forums

Work to doCompletion boundary
Find relevant real questions, then provide useful answers, demonstrations or a relevant project introduction.Check each community separately. Technical Q&A is not an ad inventory; disclose affiliation, without fabricated users or endorsements.
05 / Developer material

GitHub · Hugging Face · DEV Community · Hashnode

Work to doCompletion boundary
Real code, documentation, model information, technical tutorials and reproducible demonstrations.Use actual projects and relevant contributions, not empty repositories or unrelated promotional comments.
06 / Directories & reviews

Product Hunt · BetaList · AlternativeTo · G2 · Capterra · Trustpilot · Clutch

Work to doCompletion boundary
Maintain accurate product profiles, apply to relevant categories, and keep reviews authored by genuine users.Distinguish submission, paid application, review and publication. Account pools may manage profiles, not write fake customer reviews.
07 / Social discovery

X · Threads · Instagram · Facebook Pages · TikTok · Pinterest

Work to doCompletion boundary
Short demos, diagrams, updates, discussions and platform-appropriate product content.Check visibility per post; being published does not prove retrieval or citation by an answer engine.
08 / News & industry material

Owned newsroom · PR Newswire · Business Wire · Trade media / specialist blogs · Partner / customer sites

Work to doCompletion boundary
Genuine news, cases and primary material; offer a relevant story and evidence to editors.Paid distribution is not independent reporting. Editors control acceptance; not every outlet accepts contributed articles.
09 / Product listings

Shopify · Amazon · TikTok Shop · AliExpress · Mercado Libre · Shopee

Work to doCompletion boundary
Keep product names, variants, specifications, prices, availability, markets and media accurate.Manage authorized stores only. Listing, product recommendation, seller ranking and orders are distinct results.
10 / Private channels

Discord · Slack · Email

Work to doCompletion boundary
Collect real questions, feedback and cases; publish only authorized, redacted material.Member visibility is not public web access; do not assume private content is a citation asset.

03 / RECOMMENDATION ANALYSIS

Trace a recommendation back to a content gap.

Begin with genuine unbranded needs; test branded, comparison and purchase questions separately. Fix market, budget, use case, prompt version and time batches. Do not keep asking until the target appears.

Example: In the US, which desk arms under $150 suit two 27-inch monitors? Explain compatibility requirements.

01Recommendation

Keep the complete answer, products, ordered positions only when ordered, visible reasons and original capture/export. A mention is not necessarily a recommendation.

02Visible sources

Open actual cited pages. Separate citations, related links, verification links and retrieved candidates. Unshown sources stay unknown.

03Buying conditions

Map price, dimensions, compatibility, market, delivery and limitations to evidence. Compare what the cited page provides with what your product material lacks.

04Specific change

Turn the gap into an exact task: a specification table, product-page correction, demonstration, factual update or tailored channel content.

This is evidence inspection from visible results, not recovery of hidden reasoning. A follow-up explanation or requested score is a new answer—not the original ranking log or internal weights.

Sample ChatGPT web/shopping, OpenAI API, Gemini App/API and Google AI Overviews / AI Mode separately. Record visible model, locale, login, Memory, search and session conditions for every run; unknown stays unknown.

ANALYZEFixed questionsChatGPT / GeminiAnswers / CitationsClean entities / SourcesContent gapMatched retest

04 / CLEAN DATA

Clean observations before comparing them.

Preserve raw answers. Models propose classifications; rules check duplicates and fields; people review ambiguous products, source relationships and ownership.

ObjectTreatment
Questions & runsGroup equivalent intent; separate branded and unbranded prompts. Remove duplicate imports of one response, not independent real runs with identical answers.
Brands / products / sellersSeparate brand, family, SKU/variant and seller. Merge aliases only with evidence; review ambiguity.
URLsRetain raw, tracking-cleaned and verified canonical URLs. Remove only confirmed tracking parameters; preserve variant, color and model identity.
PublishersRecord domain, author, original work, syndication and ownership. Reprints and several owned accounts are not independent endorsements.
Link rolesSeparate citations, related/verification links and retrieved candidates. Count one URL once for within-answer coverage; store citation positions separately.
Failures & missing sourcesKeep timeouts, limits and challenges separate from non-recommendations. Retain valid answers with no visible sources in the overall denominator.
Start with four tables

runs stores conditions and raw answers; answer_items stores products and visible reasons; sources stores links and publisher relationships; actions stores changes, owners, readback and experiment IDs.

OBSERVATION CONTRACT / NOT A LIVE CONNECTOR
runs: run_id, topic_id, surface, locale, session, search_observed, captured_at, raw_answer
answer_items: run_id, brand, product, variant, seller, recommendation, visible_reason
sources: run_id, raw_url, clean_url, canonical_verified, publisher, ownership, link_role
actions: topic_id, gap, destination, owner, version, readback, experiment_id

05 / CONTENT CHECKS & RETESTS

Score content gaps, not an imaginary algorithm.

First check hard conditions: budget, market, specifications, factual support and publishing permission. Then check fit, specifications, limitations, current price/availability, comparisons, demonstrations, identity and public access.

Use 0 = confirmed missing, 1 = partial, 2 = sufficient and unknown = unchecked, each with evidence. This is an operator content checklist, not ChatGPT scores or Gemini weights. Prioritize by business relevance, recurring gaps, feasibility and cost.

ResultDenominator
Mention / recommendationCount answers mentioning and recommending the predefined target product separately, divided by valid answers under matching conditions. Repeated mentions within one answer count once. Track brand mentions separately, outside this product-level example.
Visible citation rateAnswers citing the pre-defined target pages/publishers divided by all valid answers in the group. Also report a confirmed-search subset; do not mix denominators.
Source coverageReport unique URLs, domains and independent publishers separately. Related links and candidates are not citations.
Visits & businessUse actual visits, qualified leads, orders/subscriptions. Keep branded search, direct traffic and self-reported sources separate.

Predefine prompts, repeats and time batches. Change a small topic/page group, retain comparable unchanged topics and retest under matched conditions. Report sample size, variation, model updates and parallel campaigns; a single appearance or before/after change is not causal proof.

Technical roles and what currently runs

Account tools handle authorized actions. Models may assist question grouping, writing and extraction; operators verify original text, entities and citations. Visits and business outcomes come from actual analytics. The current code starts with prepared local input, keeping API observations separate from consumer-browser results.

LOCAL REVIEW / NO ACCOUNT WRITES
python3 scripts/review.py examples.json --example geo_sample
python3 scripts/review.py examples.json --example distribution

geo_sample checks a matching surface, market, language, question/version, conditions group and target product/source. It merges identical imports of one run and rejects conflicting records. Planned, observed and valid counts remain separate from failures, refusals, non-triggered and unobserved runs. It calculates mentions, recommendations, citations and a confirmed-search subset; valid uncited answers remain in the denominator.

The synthetic before input matches the example above: six planned runs, five observed, four valid; recommendations 2/4, citations 1/4, confirmed-search citations 1/2, and one duplicate import merged. After is a separate page calculation. distribution checks an existing publication ledger. Neither connects to models, collects observations, posts, normalizes URLs/product entities or verifies citations automatically. Callers supply answer/evidence references and boolean judgments. The counts are not model weights or proof of optimization impact.

18 specific channels + a general article template.

All channel names are shown below. Select a name or use arrow keys to view its method. The original dual-monitor examples remain; choose channels for a real project, not a publish-everywhere plan. Developer communities need genuine technical work, and software directories need eligible products. ARTICLE is a general template, outside the 18-channel count. These are operating examples, not publication or citation results.

Let the image make the problem clear first.

IMAGE
A full-desk scene, an installation close-up, and horizontal/vertical screen arrangements; match every image to the facts.
COPY
Open with “check before you buy,” then list ports, weight, and desk requirements without inventing first-hand experience.
HANDOFF
Topic, image files, copy version, brand-relationship disclosure, and target account role.
CHECK
The actual images, copy, account identity, and public state; clicking Submit is not completion.

The Agent prepares the carousel and copy from verified facts. The authorized brand-account owner approves publication; reopen the post to check media, identity and public visibility. Track saves, enquiries, referrals and actual citations separately—not engagement as a citation weight.

Open the task ledger and measurement method
FieldWhy it is retained
task_id / topic / sourceTrace the task back to the original product and topic evidence
account_role / platform / content_versionKnow whom the account represents, where it was posted, and which version was used; never expose account credentials publicly
state / submitted_at / checked_atKeep PREPARED, BLOCKED, SUBMITTED_UNVERIFIED, and PUBLIC_VERIFIED separate
canonical_url / visible_copy / visible_mediaMove to confirmed only after checking the visible copy, images, and exact public URL
failure_reason / owner / next_actionLet the next operator continue; unverified work must not be blindly reposted

Published, crawlable, indexed, cited by AI, referred visits, and purchases are distinct stages. Hold the question, market, language, and review time constant, save the page actually cited, then connect visits and orders. This page currently shows the account execution and verification process; AI citations, referred visits, and orders for this content still need to be tracked separately.