Google Images vs Lens, Multisearch, and AI Mode: A Repeatable Visual Search Test | Agent Lab Journal
Agent Lab Journal
Guides
Glossary
Practical guide · Beginner level
How to Search by Image: Comparing Google Images, Lens, Multisearch, and AI Mode
35-minute read
Beginner
Repeatable four-method test
“Search by image” is not one task. Finding the original publication, naming an unfamiliar object, and answering a question about a scene require different evidence. This guide gives you one controlled image set, one scoring system, and one results table for comparing Google Images, Google Lens, Multisearch, and AI Mode without confusing a confident answer with a correct one.
Contents
Why the four modes produce different answers
What each method is good at
The concrete case
Build one fair image set
Create the answer key
Run the four searches
Score accuracy and time
Results tables
Verify claims and sources
Failure cases and recovery
Limitations
Choose the right mode
Why the four modes produce different answers
Visual search uses an image, a selected region, or an image-plus-text prompt as input. The phrase covers several different jobs:
Source discovery: locate the page, creator, publication, or earliest verifiable appearance of an image.
Object recognition: identify a product, plant, landmark, tool, artwork, animal, or other visible subject.
Scene reasoning: answer a question whose solution depends on relationships inside the image.
A tool may be strong at one job and weak at another. A visually similar page is not necessarily the source. A plausible object label may be too broad. A fluent scene explanation can contain details that are not visible.
Reverse image search is the source-oriented form of visual search: it looks for identical, altered, or visually related images already indexed on the web. Object recognition instead maps visible features to a category or entity. Scene reasoning combines visual evidence with a natural-language question.
The key rule
Judge every result against the task you actually assigned. Do not award source-discovery points merely because a result resembles the input. Do not award scene-answer points merely because the prose sounds reasonable.
What each method is designed to do
1. Google Images
Google Images is an image-results interface centered on web pages and pictures. Depending on the device and interface available to you, uploading an image may hand the query to Lens. For this test, count Google Images as a distinct method only when you begin from its image-search interface and evaluate the image results and linked pages it exposes.
It is most relevant when you need:
pages containing a matching or similar image;
alternate sizes and crops;
search terms suggested by visually similar material;
a broad survey of how an image appears across the web.
2. Google Lens
Google Lens analyzes an uploaded image, pasted image, camera view, or selected region. It is useful when the important evidence is only one part of a crowded frame. Lens may expose visual matches, product candidates, text extraction, translation, and entity suggestions.
It is most relevant when you need:
an object or landmark name;
results for a tightly selected region;
optical character recognition (OCR) for signs, labels, or packaging;
matching products or visually related objects.
3. Multisearch
Multisearch combines an image with a short text refinement. The image supplies visual context; the added words specify what you want to know or how the object should differ.
Useful refinements include:
“model name”
“replacement part”
“building architect”
“same jacket in blue”
“what is this connector used for?”
Multisearch should not be treated as an entirely separate image index. Its value is the ability to steer the visual query. Test it separately because that steering can materially change both accuracy and time.
4. AI Mode
AI Mode is a conversational search experience that can synthesize an answer from a question and available search context. Visual input availability, controls, and interface behavior may vary by account, device, region, and date. Record the exact interface you used rather than assuming every tester sees the same product.
It is most relevant when you need:
an explanation rather than a list of matches;
a multi-part answer about a scene;
follow-up questions;
-
a summary that must still be checked against visible evidence and linked pages.
Expected strengths to test—not assumed results Method Source discovery Object recognition Scene questions Main risk Google Images Candidate pages and matching images Broad visual categories Limited unless the scene is well described online Similarity mistaken for provenance Lens Visual matches, especially after cropping Objects, products, landmarks, and text Some direct answers, depending on interface Wrong crop or over-broad label Multisearch Source search with a text constraint Refined identification Focused natural-language intent Prompt wording changes the task AI Mode Possible source synthesis when links are supplied Explained identification Multi-step answers and follow-ups Unsupported or overconfident claims
A concrete case: one photograph, three different jobs
Imagine that you have a street photograph showing a bicycle beside a decorated building. A small shop sign is visible in the background. You do not know where the image came from.
The same file can generate three valid but different questions:
Source: “Which page first published this exact photograph?”
Object: “What model or type of bicycle is shown?”
Scene: “What evidence in the image suggests that the building is being used commercially?”
A page containing a repost could answer the object question correctly while failing the source question. A model could infer that the building is a shop from the sign, but it should not claim the business name if the text is unreadable. The bicycle may be identifiable only to a category, not an exact model.
This distinction determines the test design. Every image in the dataset needs a declared task and an answer key appropriate to that task.
Build one fair image set
Use the same files for all four methods. A practical starter set contains 24 images: eight for source discovery, eight for object recognition, and eight for scene questions. If that is too much for a first run, use 12 images with four per task.
Recommended 24-image structure
IDs
Task
Suggested image types
Known answer required
S01–S08
Source discovery
Original, crop, resize, screenshot, compressed copy, text overlay, color adjustment, partial obstruction
Canonical publication or documented original under your control
O01–O08
Object recognition
Common object, product, landmark, plant, tool, artwork, component, difficult angle
Verified identity at an appropriate specificity
Q01–Q08
Scene question
Spatial relationship, count, visible text, activity, safety issue, sequence clue, comparison, ambiguity
Answer supported directly by visible evidence
Use images you are allowed to test
The safest source-discovery images are your own photographs or images for which you can document the original publication. This avoids guessing who created a file and gives you a reliable answer key. Do not upload private documents, faces, addresses, medical records, access badges, or confidential screenshots to a search service.
Keep image difficulty balanced
Include both clean and altered inputs. Do not give one method only full-resolution originals and another method screenshots. Every method must receive the identical test file.
For the source group, a balanced set might include:
two unchanged originals;
one tight crop;
one resized copy;
one recompressed image;
one screenshot with interface borders;
one image with a small text overlay;
one difficult transformation, such as a partial obstruction.
Name files consistently
visual-search-test/
├── images/
│ ├── S01_original.jpg
│ ├── S02_crop.jpg
│ ├── O01_object.jpg
│ └── Q01_scene.jpg
├── answer-key.csv
├── runs.csv
└── notes/
Before testing, calculate a cryptographic hash for every file. A hash proves which exact file was used; it does not establish authorship or detect visually similar copies.
cd visual-search-test
shasum -a 256 images/* > image-hashes.txt
On systems that provide sha256sum instead:
sha256sum images/* > image-hashes.txt
If ImageMagick is installed, record dimensions and formats:
magick identify images/* > image-properties.txt
Do not repeatedly edit the test files after hashing them. If a file changes, assign a new version and calculate a new hash.
Remove accidental clues only when the test requires it
EXIF metadata can contain dates, device information, descriptions, or coordinates. Decide whether metadata is part of the challenge. For a pure pixel-based test, work on copies and inspect their metadata:
exiftool images/S01_original.jpg
If you intentionally remove metadata, document that transformation in the answer key. Never alter your only copy of an original file.
Create the answer key before searching
Ground truth is the verified answer used to judge a system. Write it before running any search. Otherwise, an attractive result can quietly redefine what you consider correct.
Create answer-key.csv with these columns:
image_id,filename,task,question,exact_answer,acceptable_answer,canonical_source,required_evidence,difficulty,notes
S01,S01_original.jpg,source,"Find the original publication",,,"YOUR VERIFIED URL","Exact image and publication record",easy,
O01,O01_object.jpg,object,"Identify the object",,"ACCEPTABLE CATEGORY",,"Visible diagnostic features",medium,
Q01,Q01_scene.jpg,scene,"YOUR SCENE QUESTION","VERIFIED ANSWER",,,"Visible region supporting the answer",medium,
Replace every uppercase placeholder with information you have independently verified. Leave fields blank when they do not apply.
Define the required specificity
“Bicycle” may be correct at category level but wrong when the task asks for a specific model. Conversely, an exact model should not be required when the image lacks diagnostic details. Record the narrowest answer that the image and your evidence can support.
Use three answer bands:
Exact: the expected source, entity, or complete scene answer.
Acceptable: correct but less specific, with no contradictory claim.
Incorrect: wrong, unsupported, or missing the requested information.
Prepare source-discovery evidence
A canonical source should be supported by records you control or can verify, such as the original file, publication date, creator record, or a documented upload. Search ranking alone cannot prove that a page was first.
For public images whose true first appearance is uncertain, change the task from “find the original source” to “find the known canonical page.” This narrower claim is testable.
Run the four searches under the same conditions
1. Freeze the test conditions
Record the following before the first query:
date and local time;
country or region;
device and operating system;
browser or app;
signed-in or signed-out state;
interface language;
whether search history or personalization is active;
network type;
which visual-input controls are available.
Search indexes and interfaces change. These details make your result interpretable without pretending it is timeless.
2. Define when the timer starts and stops
Use a stopwatch. Start when you submit the uploaded image or image-plus-text query. Stop at the first moment the required answer is visible and sufficiently supported for the task.
If you must open a result page to confirm the source, the timer continues. If a candidate appears quickly but verification takes longer, record both values:
candidate time: first plausible answer;
verified time: sufficient evidence for scoring.
Set a fixed limit, such as five minutes per image per method. A time limit is not a claim about normal performance; it is a rule for your experiment. Mark unresolved runs as timeouts instead of searching indefinitely.
3. Prevent learning from one method affecting another
Once you learn an object name from the first method, you may unconsciously use it in later queries. Reduce this carryover by rotating method order:
Image sequence
Method order
1, 5, 9, 13, 17, 21
Images → Lens → Multisearch → AI Mode
2, 6, 10, 14, 18, 22
Lens → Multisearch → AI Mode → Images
3, 7, 11, 15, 19, 23
Multisearch → AI Mode → Images → Lens
4, 8, 12, 16, 20, 24
AI Mode → Images → Lens → Multisearch
Do not paste a label learned from another method unless the protocol explicitly allows follow-up searching. Keep raw-run and assisted-run results separate.
4. Use fixed query rules
Google Images procedure
Open the image-search interface available on your device.
Upload the unchanged test file.
Do not add text during the initial run.
For source tasks, inspect exact or near-exact matches and open candidate pages.
For object and scene tasks, record the best answer visible within the time limit.
-
Save the result URL when possible and note any redirection into Lens.
Lens procedure Upload the identical test file.
Begin with the full image.
If no result answers the task, make one crop adjustment only.
Record whether the successful result came from the full frame or selected region.
-
Do not type an object name learned elsewhere.
Multisearch procedure Upload the identical test file.
Use one predefined text refinement based only on the task.
Do not include the answer or an unverified guess in the refinement.
Record the exact text, including capitalization and language.
Allow one predefined fallback refinement if every method receives the same opportunity.
Recommended neutral refinements:
Task
Initial refinement
Optional fixed fallback
Source discovery
original source
page containing this exact image
Object recognition
exact name or model
identify using visible features
Scene question
The exact question from the answer key
answer only from visible evidence
AI Mode procedure
Confirm that the interface permits the same image input. If it does not, record “not available” rather than substituting another workflow.
Attach the identical file.
Use the exact question from the answer key.
Add: State uncertainty and identify the visible evidence for your answer.
For source tasks, require a page that can be opened and checked.
Record the initial response before asking follow-up questions.
Keep follow-up performance in a separate column. Otherwise, a multi-turn AI run is being compared unfairly with a single submission in the other modes.
5. Log every run
Create runs.csv:
run_id,image_id,method,method_order,query,crop_used,start_time,candidate_seconds,verified_seconds,outcome,score,confidence,top_result_url,evidence,error_type,notes
1,S01,Google Images,1,"image only",no,,,,,,,,,,
2,S01,Lens,2,"image only",no,,,,,,,,,,
3,S01,Multisearch,3,"original source",no,,,,,,,,,,
4,S01,AI Mode,4,"Find the original source. State uncertainty and identify the visible evidence for your answer.",no,,,,,,,,,,
Use the same spelling for every method and outcome. Consistent labels make totals easier to calculate.
Score accuracy and time
A simple three-level accuracy score
Score
Meaning
Source example
Object example
Scene example
2
Exact and supported
Verified canonical page found
Correct identity at required specificity
Complete answer supported by visible evidence
1
Useful but incomplete
Correct repost or strong lead, not the canonical page
Correct category but insufficient specificity
Partially correct with no harmful contradiction
0
Incorrect, unsupported, absent, or timed out
Similar image presented as the source
Wrong entity or unjustified exact model
Contradicted by the image or based on invented detail
For stricter reporting, publish both exact accuracy and useful-answer rate:
Exact accuracy (%) = exact answers / eligible images × 100
Useful-answer rate (%) = (exact + partial answers) / eligible images × 100
Mean score (%) = total points / (eligible images × 2) × 100
An eligible image is one for which the method was genuinely available. Do not count “AI Mode did not accept image input” as a wrong visual answer. Report availability separately.
Score task-specific evidence
Source discovery: confirm that the page contains the same image, not merely a similar subject. Compare distinctive details, crops, overlays, and orientation. Then assess whether the page satisfies the predeclared source definition.
Object recognition: compare the answer with your verified label. Penalize unjustified specificity. “Likely a compact folding bicycle” can be better than a confident but unsupported model name.
Scene questions: split the expected answer into required facts. Award full credit only when all facts are correct and supported by visible regions. A correct conclusion with fabricated supporting details is not exact.
Report median time as well as average time
The median is the middle measured value after times are sorted. It is less distorted by one unusually slow run. Report:
median candidate time;
median verified time;
average verified time;
number of timeouts;
number of unavailable runs.
Do not convert timeouts into the time limit when calculating the main median unless you explicitly declare that rule. A clearer approach is to report completed-run time and timeout count separately.
Accuracy and time results tables
The tables below are deliberately unfilled. Publishing numbers without running the protocol would create test results that do not exist. Copy the tables, complete the experiment, and replace “Not measured” with calculated values.
Primary comparison across one shared image set
Method
Eligible images
Exact accuracy
Useful-answer rate
Mean score
Median candidate time
Median verified time
Timeouts
Google Images
Not measured
Not measured
Not measured
Not measured
Not measured
Not measured
Not measured
Google Lens
Not measured
Not measured
Not measured
Not measured
Not measured
Not measured
Not measured
Multisearch
Not measured
Not measured
Not measured
Not measured
Not measured
Not measured
Not measured
AI Mode
Not measured
Not measured
Not measured
Not measured
Not measured
Not measured
Not measured
Accuracy by task
Method
Source exact accuracy
Object exact accuracy
Scene exact accuracy
Unsupported-claim count
Google Images
Not measured
Not measured
Not measured
Not measured
Google Lens
Not measured
Not measured
Not measured
Not measured
Multisearch
Not measured
Not measured
Not measured
Not measured
AI Mode
Not measured
Not measured
Not measured
Not measured
Calculation worksheet
For each method, count exact, partial, incorrect, timeout, and unavailable outcomes. Then calculate the rates only from eligible images.
eligible = exact + partial + incorrect + timeout
exact_accuracy = exact / eligible
useful_rate = (exact + partial) / eligible
mean_score = (exact * 2 + partial) / (eligible * 2)
availability = eligible / total_images
Present percentages with counts, for example x/y, so a reader can see the sample size. Avoid declaring a universal winner from a small dataset. Report the strongest method separately for source, object, and scene tasks.
Verification: how to tell whether an answer is actually correct
Verify an image source
Open the candidate page; do not score from a thumbnail alone.
Confirm that the same image appears on the page.
Compare stable details: object positions, shadows, damage, reflections, and background geometry.
Determine whether differences are transformations such as cropping, mirroring, resizing, or added text.
Check whether the page identifies a creator or links to an earlier publication.
Compare the page with the canonical source recorded before the test.
If priority cannot be established, label the result “matching publication,” not “original source.”
Verify an object
List the visible diagnostic features before reading the result in detail.
Compare those features with the proposed identity.
Look for contradictions, not only similarities.
Reduce specificity if the necessary model-level markings are absent.
Use text on labels only when it is legible in the source image or confirmed independently.
Verify a scene answer
Break the response into individual factual claims.
Mark each claim as visible, reasonably inferred, externally sourced, or unsupported.
Check counts and spatial relationships directly against the image.
Reject invented text, identities, intentions, dates, or locations.
Score the answer against the prewritten answer key, not its writing quality.
Confidence is not evidence
A method’s wording—“this is,” “likely,” or “possibly”—can help communicate uncertainty, but it does not prove the claim. Verification must use the image, the answer key, and inspectable pages.
Have a second person score ambiguous runs
If possible, hide the method name and ask a second reviewer to score the captured answer. Compare disagreements and resolve them using the written rubric. This reduces the chance that a preferred interface receives more generous grading.
Common failure cases and what to do
No exact image matches
The image may be private, new, poorly indexed, heavily altered, or absent from the searchable web. Try a crop containing the most distinctive region. Record the crop as a separate attempt; do not silently replace the original run.
Only visually similar images appear
Similarity can help generate keywords but does not establish a source. Use visible landmarks, text, product markings, or unusual composition to form a follow-up query. Score the initial source run as partial or incorrect according to the rubric.
The wrong object dominates the query
A busy background can overpower a small target. Select the object more tightly while retaining diagnostic context. Record that a crop was required because crop-assisted performance is different from zero-interaction performance.
OCR returns convincing nonsense
Small, curved, reflected, stylized, or blurred text is easy to misread. Compare extracted characters with the pixels. Search several plausible readings only in a documented follow-up phase.
A model gives an exact product name without visible markings
Treat unjustified specificity as an error. Look for model numbers, port layouts, fastener positions, logos, dimensions, and other diagnostic details. If those are unavailable, accept only the broader category defined in the answer key.
AI Mode answers the question but supplies no checkable support
For a scene question, visible evidence may be enough. For a source claim or obscure identification, require an inspectable result page or independent confirmation. An unsupported answer is not exact even when it later turns out to be right.
AI Mode or Multisearch is unavailable
Record the interface, account state, region, device, and date. Mark the run unavailable. Do not imitate the feature by manually combining outputs from other modes unless you define that as a fifth workflow.
The interface changes during the test
Pause the run and document the change with notes or screenshots that contain no private information. Resume only if the same workflow remains possible. Otherwise, divide the experiment into versions.
Personalization may affect results
Repeat a small subset in a fresh browser profile or signed-out session when feasible. Treat this as a sensitivity check, not a guaranteed removal of every location or device effect.
The “original” page is newer than a repost
Publication dates can be edited, imported, or displayed without enough context. Do not infer authorship solely from the oldest visible date. Fall back to the narrower, defensible label “earliest verified page in this test.”
Limitations of this comparison
Search results change. Indexes, rankings, interfaces, and generated answers are not fixed.
Availability varies. A feature may differ by account, language, region, device, or application version.
The dataset shapes the winner. Product-heavy images may favor one workflow; provenance-heavy images may favor another.
A small set has high uncertainty. Twenty-four images can reveal workflow differences but cannot represent the entire web.
Timing includes human work. Reading speed, crop choices, and verification skill affect measured time.
Source discovery has an open-world problem. The true first publication may be deleted, private, unindexed, or otherwise inaccessible.
Prompting affects conversational modes. Small wording changes can alter the answer. Preserve exact prompts.
Image quality is a confounder. Resolution, compression, glare, orientation, and occlusion influence results.
Correctness can require expertise. Plants, medical imagery, industrial components, and artworks may need a qualified reviewer.
Do not use a general visual-search result as the sole basis for medical, legal, safety, identity, or authentication decisions. Escalate consequential conclusions to an appropriate expert and preserve the original evidence.
Which mode should you try first?
Your immediate goal
Start with
Then try
Do not confuse with success
Find pages containing the same image
Google Images or Lens visual matches
Crop distinctive regions; add a neutral source refinement
A similar image of the same subject
Identify one object in a busy frame
Lens with a careful region selection
Multisearch with “exact name or model”
An unsupported model-level guess
Read a label or sign
Lens OCR
Search the verified characters with object context
Plausible text not visible in the pixels
Ask a multi-part question about a scene
AI Mode, if visual input is available
Lens or Multisearch for individual objects and text
A fluent explanation containing invisible details
Establish provenance
Matching-image pages plus manual verification
Creator records and known canonical publication data
The oldest date displayed in one result
A practical combined workflow
State whether you need a source, an identity, or a scene answer.
Upload the full image to obtain an initial overview.
Select the most distinctive relevant region.
Add a neutral text refinement only after preserving the initial result.
Use conversational synthesis for explanation, not as proof.
Open candidate pages and verify every material claim.
Record accuracy, candidate time, verified time, and failure type.
The most useful outcome is rarely “one tool wins.” A defensible conclusion looks more like this: one method was fastest for locating candidate pages on this dataset, another produced the most exact object labels, and another made scene questions easier to explore but required more claim checking.
Repeatable test checklist
Use 12 or 24 images with known answers.
Divide them evenly among source, object, and scene tasks.
Hash the exact files and preserve their properties.
Write the answer key before searching.
Use identical files and fixed query rules.
Rotate method order across images.
Record candidate and verified time separately.
Apply the same 0–2 rubric to every method.
Keep initial and follow-up performance separate.
Mark unavailable features instead of inventing substitutes.
Verify sources by opening pages.
Publish counts, percentages, timeouts, and limitations.
Next: continue with the practical material in Agent Lab Journal guides, or review unfamiliar concepts in the glossary.
Agent Lab Journal
Top comments (0)