Can Writesonic summarize a PDF with tables and keep formatting?

I have been conducting a series of methodical evaluations on AI-powered document processing tools, specifically focusing on their ability to handle complex, data-dense PDFs. My primary use case involves technical whitepapers, financial reports, and research documents where tabular data is not merely supplemental but integral to the core argument. The central question for this thread is whether Writesonic's document summarization can accurately preserve and interpret table structures, or if it degrades to plain text, losing relational context.

My testing methodology involved three distinct PDF document types, each with increasing table complexity:

1. **Simple Table:** A document with a basic 3x4 table showing quarterly sales figures (numerical data with clear headers).
2. **Nested Table:** A research paper containing a multi-header table with merged cells, comparing database performance metrics (latency, throughput, connection limits).
3. **Formatted Financial Statement:** A PDF with embedded stylized tables, including row shading, currency symbols, and column-spanning subtotals.

I used Writesonic's "Chatsonic" interface with the "Summarize Document" capability, uploading each PDF directly. The following are my raw observations on formatting preservation:

* **Text Extraction & Basic Summary:** The engine reliably extracts the textual content from cells. A summary of the surrounding paragraphs is generated competently.
* **Tabular Data Handling:** The system identifies the *presence* of tabular data, often prefacing its summary with statements like "The document contains a table showing..." However, the actual table data is then presented in a linear, unstructured format.
* For the **Simple Table**, data points were listed in a sentence-like structure: "Q1 sales were $X, Q2 sales were $Y..."
* For the **Nested Table**, the hierarchical relationships were lost. Column headers and row contexts were often conflated, making the output misleading for detailed comparison.
* The **Formatted Financial Statement** lost all visual formatting cues. Subtotals were not clearly distinguished, and the structural meaning implied by indentation and grouping was absent.

Crucially, the output lacked any markdown table syntax (`| --- |`) or structured data formats (JSON, CSV). This suggests the summarization pipeline is operating on a pre-processed plain-text representation of the PDF, rather than applying a layout-aware parsing model that understands cells as discrete, related entities.

For comparison, I performed a similar test using a custom script with the `pdfplumber` library for extraction and a leading LLM API for summarization with a specific prompt to "retain table structure." The difference was stark, as the LLM could be instructed to output in markdown.

```python
# Example of a prompt that yields structured output from raw extracted text
prompt = f"""
Summarize the key data from the following document excerpt.
CRITICAL: Preserve any tabular data exactly in markdown table format.

Document Text:
{extracted_text}
"""
```

**Conclusion:** If your requirement is a high-level, gist-based summary of a document's narrative, Writesonic performs adequately. However, if accurate preservation of tabular formatting and relational data is necessary for your workflow—such as for data validation, comparative analysis, or feeding summarized tables into another system—the tool in its current implementation is insufficient. The loss of structure introduces an unacceptable risk of misrepresentation for analytical purposes. I am interested in whether other users have developed prompt engineering techniques or pre-processing workflows to mitigate this limitation, or if alternative document-aware AI tools have proven more capable in this specific niche.