TL;DR
- Use tables when they make a comparison clear. Check what each row and column means, as well as the layout.
- Older tests found different success rates for different tasks on the same HTML format. They are not current AI-search error rates.
- Text-similarity scores in another study are not the share of claims that are true.
- Check the product, plan, unit and time basis behind an AI summary. Make missing values and exclusions clear.
Structure Helps, but the Answer Still Needs Evidence
Microsoft Research’s March 2024 article asks how well language models understand tables in prompts. Its useful starting point is that results depend on more than having a table. A structured layout and a correct reading are different questions.
For generative engine optimization (GEO), this gives a content team a clear task. A buyer should be able to trace “this plan includes the full allowance” to the right cell and note. A neat table with an unclear unit has not solved that task.
The Same HTML Format Had Different Task Results
Sui and colleagues’ WSDM 2024 paper reports HTML-input accuracy of 44% for cell lookup, 42% for row retrieval, and 67% for size detection. These are from its default text-davinci-003 setting. The one-shot test used Wikipedia-derived tables. It sampled 1,500 tables per task setting after filtering out easy questions.
These are older lab results, not a test of current web-search tools. “The model can read HTML” is too broad a completion check. Finding where a value sits, returning a row and counting table dimensions are different tasks.
Sui et al., WSDM 2024, Table 2, default Acc columns; selected tasks. The benchmark samples 1,500 tables per task after difficulty filtering. These are separate tests, not a funnel, a current model ranking or AI citation rates. Original paper, Table 2 and methods.
A Good Text Score Is Not a Fact-Checking Pass
Borisova and colleagues’ 2025 TableEval study used 3,017 tables from scientific papers, Wikipedia and financial reports. Across its models and formats, mean ROUGE-L.F1 was 0.36 for non-scientific subsets and 0.15 for scientific subsets. This score compares wording with reference text. It is not the share of product claims that are true.
The domains also have different tasks and reference styles. The comparison does not isolate an effect of being scientific. Check the question and scoring rule before applying a benchmark result to your own table.
| Direct research | Actual result | What the score tests |
|---|---|---|
| Sui et al., 2024 | HTML: cell lookup 44%, row retrieval 42%, size detection 67% | Structural task accuracy in one historical prompt setup |
| Borisova et al., 2025 | ROUGE-L.F1: non-scientific 0.36; scientific 0.15 | Reference-text overlap, averaged across models and formats |
Figure 2. Evidence comparison. Different teams, datasets and outcomes. A structural accuracy and a text-overlap score cannot be averaged into a shared reliability rate. Neither study tests live B2B citation frequency.
This does not mean “replace every table.” Keep a table when it helps compare like with like. Add the context a buyer needs to read each cell. Then check the answer against the facts. A format cannot certify a claim.
A Completed Check of a Plan-Table Summary
This example is fictional. A software page compares two plans. The approved table says Standard includes 500 file exports per account per calendar month. Pro includes 1,000. Each file export counts once. Extra exports are not included. A hypothetical AI summary says, “Standard includes 500 exports per user, with extra exports included.”
| Claim to check | Approved fictional fact | Review result |
|---|---|---|
| Which plan has the allowance? | Standard: 500; Pro: 1,000 | The summary used the Standard value |
| What is the unit and period? | File exports per calendar month, shared across the account | “Per user” is unsupported; keep account scope beside the value |
| Are extra exports included? | No; extra exports are not part of the stated allowance | The summary reverses the exclusion |
| What should the corrected answer say? | Standard includes 500 file exports per account per calendar month; extra exports are not included | Check the full corrected sentence against the approved source |
The number 500 was copied correctly. The answer was still wrong about scope and exclusions. Checking just the numbers would miss a key business error. Check the value and its meaning together.
On the public page, put the unit and period in the heading or near the value. Name each plan. Keep exclusions where a reader can link them to the right cell. Do not use a blank cell to mean both “not included” and “not supplied.” Those states mean different things.
If a table has been merged, flattened or taken from a document, check that each value still has the right row and column labels. Inspect the version the buyer or tool gets. A screenshot of the old layout cannot prove what a later summary used.
These are editorial checks, not a tested way to gain AI citations. Our PDF and HTML guide covers access and formats. This article asks how to keep the meaning intact once the facts are presented.
AI Comparison-Table Answers: Common Questions
Should we replace tables with prose?
Not as a blanket rule. Use the form that makes the buyer’s comparison clear. Explain conditions that do not fit neatly in cells. A paragraph does not itself prove that an AI answer will be correct.
Does valid HTML prove an AI understood a table?
No. Valid markup can describe the structure, but it does not verify a claim. Check the product, row, column, unit and condition in the output. A syntax check and a meaning check do different jobs.
Can a high text-similarity score replace manual fact checks?
Not for that claim. A summary may share many words with a reference while changing a negation or plan scope. A score can help compare outputs in a defined test. The product fact still needs its own check.
What should a content team count as a completed repair?
The approved fact and its conditions should be clear on the final page. Check any saved wrong answer against that source. A repaired page is a content result. Whether an AI later uses it correctly is a separate observed outcome.
Bring the Table and the Claim It Produced
If buyers keep seeing the right number with the wrong meaning, bring the table, approved plan facts and saved AI summary to SemShift’s GEO service discussion. We can discuss which headers, notes or source passages leave room for error and what a clear repair would need to show.