Case 15
∅ No visible movement
repeated-prompt log How can a business test one optimisation change
Read → A practical research pattern for testing one LLM optimisation change without mixing levers or overstating what the answer proves.
Research notes
A corpus of lever tests
Levertrace Lab publishes research notes on how LLMs describe French businesses after public optimisation changes. The materials are organised by lever type, answer behaviour and language condition, so readers can compare owned-page edits, directory corrections, structured signals and bilingual source alignment without treating them as one blurry tactic. New notes are added as the team completes enough comparison to say something cautious and useful.
Lever tests · 3 directions · English, French
Direction I
Owned-page & structured levers
Changes a business controls directly: owned-page edits, category wording, structured data and dated updates. The lab traces whether a model adopts the new wording, echoes part of it, or leaves the old description in place.
Case 10
● Adoption
before/after test Does site structure affect LLM extraction
Read → How Levertrace Lab studies whether clearer linking, hierarchy and entity cues help models describe a French business less vaguely.
Case 03
∅ No visible movement
before/after test Does structured data change the LLM answer
Read → An investigation into whether schema markup alters LLM business descriptions or mainly supports systems that read structured search signals.
Case 02
● Adoption
before/after test Which wording change moves the business category first
Read → A focused study of category wording, service boundaries and entity phrasing as single-page levers in LLM business descriptions.
Case 01
~ Partial echo
before/after test Does editing the website change the LLM description
Read → A close look at when owned-page edits appear in LLM descriptions of French businesses, and why movement is often delayed or partial.
Direction II
Third-party & source levers
Directory corrections, trade pages, review platforms and business profiles competing with the company’s own site. The lab watches for the third-party source that quietly holds the model’s wording, even after the owned page has changed.
Case 13
⚇ Source conflict
source comparison Which levers appear only in live search
Read → A field note on why some LLM optimisation changes appear in live-search answers but stay absent from memory-based descriptions.
Case 09
● Adoption
source comparison Which French platforms change LLM output
Read → A source-lever study of French directories and business platforms, asking which updates appear in model descriptions and which do not.
Case 05
⚇ Source conflict
source comparison Does repeated source agreement outweigh one cleaner source?
Read → A study of whether LLMs follow many matching public claims or one clearer source when describing a French business.
Case 04
● Adoption
source comparison Does a corrected directory listing move faster than the company site?
Read → A close look at whether third-party corrections can alter LLM business descriptions before owned-site edits do.
Case 07
⚇ Source conflict
source comparison Do owned pages or earned mentions shape the summary
Read → A close look at whether company-controlled pages or earned public mentions seem to guide LLM summaries of French businesses.
Direction III
Language & persistence levers
Bilingual rewrites, interlinking and repeated claims read across French and English. The lab follows whether a French correction propagates into English answers, distorts on the way, or leaves an outdated fact stuck in place.
Case 12
~ Partial echo
before/after test Does fresh dated content change the answer
Read → Levertrace Lab examines whether dated updates and refreshed pages help newer French business facts appear in LLM descriptions.
Case 14
∅ No visible movement
repeated-prompt log Does the same lever work across models
Read → A comparison note on why one optimisation change may alter ChatGPT, Claude, Gemini or Mistral answers unevenly for a French business.
Case 11
● Adoption
repeated-prompt log Do repeated claims increase stable LLM wording
Read → A focused study of whether consistent category and capability claims across pages make model descriptions of French businesses steadier.
Case 06
~ Partial echo
language-pair check Does a French edit reach the English answer?
Read → A study of whether French source corrections propagate into English LLM descriptions, and when English rewrites distort the French facts.
Case 08
∅ No visible movement
repeated-prompt log Is fixing a wrong fact harder than adding one
Read → A before-and-after look at why LLMs may adopt new missing facts faster than they abandon old incorrect ones.
Each research note follows a lever from edit to answer wording.
Browse the cases by lever type, answer behaviour and language condition.
About the method →