Some e-commerce merchants’ customers don’t know how to use the product and end up leaving negative reviews. Damon Chen heard about this problem while running Testimonial. Testimonial helps merchants collect video testimonials; merchants wanted to raise their collection rate but found that customers did not even understand how to use the product. The direction Damon later tried was to let customers ask the product manual directly. Manuals are usually in PDF. [1]
This need later grew into PDF.ai. The first direction explored was to let merchants’ customers chat with product manuals; today, what has official documentation available is its API. Developers submit a PDF, have it broken into readable content, extract fields according to a schema they write, or ask questions against an already parsed document. The API is called by integrators holding a key, who buy credits. The API returns parsed content, structured fields, and question-and-answer answers, but it does not write the schema for the integrator, nor does it put the answers back on the customer service page.
Image: Source PDF.ai official website[2], retrieved 2026-11-10, original interface preserved.
Parsing is a foundational step, but not every flow has to call Parse separately first: Extract can use file/url to parse along the way in the extraction call and return a docId, while Ask requires that the docIds have already been parsed. The standalone Parse endpoint can upload a file, provide a URL, or take an existing docId and return readable content, page count, and docId. standard uses OCR and costs 1 credit per page; advanced uses a vision-language model and costs 2 credits per page; if an LLM is made to process images, add 1 credit for every 5 images, rounded up. [3]
The docId is the discount key on the bill. Calling Parse again with the same docId under the same settings hits the cache and is charged 0 credits; this docId can also be reused for later extraction and splitting calls, avoiding repeated parsing. What is saved is repeated PDF parsing, not every subsequent call.
Extract requires you to write a JSON schema first; it returns data according to the schema and attaches citations: each has content, pageNumber, and schemaLink, pointing the field back to the original passage and page number. It does not just give a summary; it hands over “where this number comes from” as well. But the schema is written by a person; if the field definition is wrong, no matter how accurate the citation is, it is only accurately pointing to a wrong question. [4]
Ask accepts a prompt and one or more docIds; the docIds passed to it must already have been parsed. credits are charged at the total page count of all documents in the request ×3. For example, a 200-page product manual costs 600 credits per question, before counting parsing costs. The thicker the manual, the more expensive each question, regardless of how detailed your question is. [5]
The free plan for the API is $0/month and includes 200 credits/month, with no credit card required; paid plans start at $49/month for 3,000 credits and increase by tier, and above 100,000 credits/month you need to contact them. [6]
For Extract billing, look at the endpoint formula first; don’t be led astray by the short summary on the credit usage page. The endpoint documentation calculates by the number of schema fields: 2 credits per page for 5 or fewer, and 4 per page for more than 5; parsing fees are separate—passing file/url generates parsing credits in the extraction call, while passing an existing docId does not. The official credit usage page’s short summary says standard is 2 per page and advanced is 4 per page, with different labels; the official billing note is explicit that when details conflict, the endpoint page prevails, and Extract calls should be estimated by field count. [4] Billing note[7]
Damon did not first want to build an AI PDF tool and then look for a use case. In 2020, he was still an engineer at Cisco, and his child was one year old. He applied for 6 months of unpaid leave to work on side projects; he made 5, and the first 4 brought in $0 in revenue; the 5th, Testimonial, reused about 80% of the code from Indielog. When the leave ended, he returned to Cisco and kept working nights and weekends; 4 months later, he quit after reaching $2K MRR. Quitting cut his household income by 50%; he and his wife agreed that if he did not reach $100K ARR within one year, he would return to a tech company, and during that period his wife covered the mortgage and bills. [1]
PDF.ai grew out of the merchant need behind Testimonial: first there was the specific trouble of customers not knowing how to use the product and leaving negative reviews; only then came the attempt to let customers chat with the product manual. This order explains why it is not simply “upload a PDF and chat”: what merchants want is to put the answer back in front of the customer, while the interface breaks this into three parts—parsing, extraction, and asking.
The interview introduction says PDF.ai is over $500K ARR and Testimonial is over $800K ARR; this is a self-reported business statement at the time of the interview and cannot be treated as the current revenue of PDF.ai. How much of the business is the API, and how many paying customers there are, are not disclosed in the public version; the interview cuts off where it asks about acquisition channels. Public materials also do not include inference and OCR costs, file storage, or customer service expenses, and profit cannot be inferred from credits pricing. [1]
The comparable player in the same category is ChatPDF. Its backend API also first adds a PDF source, with a single file limit of 2,000 pages or 32 MB; the chat endpoint is stateless, so history must be included in one request, with a maximum of 6 messages, and when the total token count exceeds 2,500, earlier messages are ignored; with referenceSources enabled, the response can include inline page numbers in the form [P<N>] and a references array. By contrast, Ask from PDF.ai requires docIds to have already been parsed, charges credits by the total page count of all documents ×3, and can pass multiple docIds at once. What can be compared side by side is only the interface conditions, not answer accuracy on the same document. [8]
Back to that kind of bad review from the beginning: what merchants first want to solve is customers not knowing how to use the product and leaving negative reviews, not just “AI can read PDFs.” PDF.ai breaks the manual into Markdown, fields, and page citations, and it also breaks this into credits deducted by page. Caching the docId avoids repeated parsing, but it does not avoid extraction being billed by page or asking being billed by the whole book’s page count. The interface can find the answer for the merchant, but it cannot judge for the merchant whether the schema is written correctly, nor can it make the customer ask the question clearly. What this API really sells is not reading a PDF once, but turning “the customer cannot find the answer” into a budgetable call.