Your Spec Sheet Is a PDF. AI Reads None of It.

A precision hardware factory in Shenzhen publishes one line on its product page: download the spec sheet for full details. A buyer downloads the 12-page PDF, then turns to ChatGPT and asks whether this shop can hold ±0.01 mm. The answer comes back: contact the manufacturer to confirm.

The buyer never contacts them. He goes to another supplier, one whose tolerances sit in a plain HTML table that AI reads in a single pass.

So in AI search optimization, the question about a PDF is not whether you have one. It is whether the sentence inside can be lifted out on its own. A file existing is not the same as text existing.

Boundaries first. This piece is about technical documents — spec sheets, drawings, inspection reports, parameter tables — and how to organize them so AI can read them and quote them. It is not about PDF layout or design, and it does not promise that turning a PDF into a web page will produce inquiries. If your buyers never read technical documents at all, because you sell consumer goods, most of this will not apply. One last limit: GEO can make parameters that already exist readable. It cannot invent data you never measured.

Google does index PDFs, but through an extractor

Many people assume a PDF is invisible to search. It is not.

In Google's own documentation, "File types indexable by Google," PDF sits in the list of supported encoded file types. The same table lists Excel (.xls, .xlsx), Word (.doc, .docx), PowerPoint (.ppt, .pptx), PostScript and EPUB.

But the same paragraph carries a qualifier: these encoded formats are, in the documentation's words, binary files or complex containers that require a specific parser to extract human-readable text.

That sentence decides what kind of problem this is. A PDF becomes readable not because the file exists, but because a parser can pull text out of it. If it can, you have text. If it cannot, you have nothing.

In generative engine optimization the difference from a web page is structural. On a page the words are already inside the HTML: fetch it, and you have text. A PDF has a shell to open first. If the shell holds characters, fine. If it holds something else, the chain breaks exactly there.

Three break points, from common to fatal

To judge a PDF you only need one question: are those characters real text, or things that merely look like text?

First break point: the scan. Inspection reports, certificates, quality records — many factories produce these by scanning paper. The whole PDF is a single image and not one character is encoded. The parser extracts nothing, so AI sees a picture rather than a paragraph.

Second break point: drawings. A CAD export turns dimensions, tolerances and material notes into line work and text blocks. Some of those letters are paths rather than characters: on screen you read ±0.01, but the file contains no encoding for those glyphs. There is a direct test. Open the PDF and try to select that line with your cursor. If it selects, it is text. If it will not select, it is a path.

Third break point: tables. A parameter table is the most fragile structure inside a PDF. The text comes out, the row and column relationships do not. Type A at 0.5 mm and Type B at 1.2 mm get flattened into one line, and AI can no longer tell which value belongs to which type.

Of the three, only the third can be rescued. The first two are cheaper to redo than to repair: re-export a text-layer version from the source system. Which gives a working rule for GEO: do not begin with how do we optimize this PDF. Begin with does this PDF contain extractable text at all.

Why factories love PDFs

It is not laziness. A PDF has three genuine advantages: the layout is locked, so what a buyer prints matches what you designed; the file travels well, and a 20-page spec sheet survives email intact; and a technical document reads as official in a way a web page does not.

The habit comes from the old export workflow. Quote, sample, spec sheet — every step was point to point. The buyer asked, you sent. Whoever held the file held the information.

That workflow made sense while the first filter was a trade fair or a referral. A buyer could not compare tolerances across forty suppliers without asking each one, so the PDF was the pitch. Once AI search became a purchasing entry point, the first step changed: the buyer no longer asks you first, he asks AI whether the specification can be made at all.

At that step the PDF is invisible. It is not in the answer, because it is not a paragraph that can be quoted.

One figure is worth keeping in view. In AirOps's State of AI Search 2026, 85% of the brand mentions found in AI answers originate on third-party pages, and 83% of commercial citations point at pages updated within the previous year. Both are vendor-published observational statistics rather than experimental results. Both point at the same thing: AI takes text blocks it can lift, not files.

That explains a shared confusion among factories in Shenzhen and Dongguan: the material is complete, yet AI cannot find them. Usually the material is not the problem. The carrier is.

The fix is three layers, not a binary choice

The usual advice is to move the PDF content onto a web page. That is half right. Move all of it and the page becomes a technical manual nobody reads to the end. Move none of it and AI reads nothing.

A workable structure has three layers.

Layer one: a summary table on the page. Not every parameter — the fields a buyer needs in order to decide: material, size range, tolerance, surface finish, capacity, lead time, certification. Written as an HTML table, never as an image. This layer exists for AI to read and for the buyer to skim.

Choosing those fields is easier than it sounds. Pull the last fifty inquiry emails and mark every question that contains a number or a unit. The fields that repeat are your table. Anything asked once stays in the PDF. Six to ten fields is the practical range; past that the table stops being skimmable and starts being a document.

Layer two: keep the PDF as the detailed version. Full specifications, test methods and packaging data stay inside it, with the link placed below the table. This layer serves the buyer who has already decided to go deeper. One detail earns its effort: open every PDF page with a complete sentence describing what that page covers, instead of a bare heading. That sentence is the part most likely to be lifted as a summary.

Layer three: build HTML spec pages for key products. For your main models, one pure-HTML technical page each, on its own URL, indexable and quotable on its own. This layer costs the most, so it belongs to flagship products only, not the whole catalogue.

That division of labour is also the difference between GEO and producing another brochure: layer one competes to be quoted, layer two absorbs depth, layer three builds a long-lived asset.

One instruction points the other way: do not expect an image-only PDF to be read. If your inspection report is a scan, the answer is not optimization but re-exporting a text-layer version from the original system. Rebuilding it with OCR is a bad trade, because one wrong digit in an OCR'd tolerance is worse than no tolerance at all. This is one of the few places in GEO where the right move is to redo rather than to rewrite.

Drawings can only change carrier

Drawings are the hardest category, because their information was never designed for prose.

An assembly drawing carries dimensions, tolerances, material, surface treatment, assembly order, critical fits. On the drawing these are symbols and callouts; in a conversation they are sentences. Nothing converts one into the other automatically.

So for drawings the GEO move is not making AI read the image. It is answering in words the handful of questions buyers actually ask.

For example: maximum part size 1200 x 600 mm; standard tolerance ±0.02 mm, down to ±0.01 mm on wire EDM; common materials 6061, 7075, 304; surface options anodizing, bead blasting, plating.

Those four sentences are worth more than a drawing, because the questions buyers put to AI are precisely those four sentences.

A second risk sits on top of this. When AI quotes you it drops your qualifiers, keeping the main clause and shedding the adverbial ones. So can hold ±0.01 mm, requires wire EDM, roughly 30% added per part is more useful than high precision; and you must accept that conditions inside brackets will sometimes be lost. That is why critical limits deserve their own sentence rather than a parenthesis, and why writing for AI search optimization differs from writing a catalogue.

What to measure afterwards

Do not fixate on whether the PDF got indexed. Look at something else: when a buyer asks a parameter question, does the answer contain the specific values from your table?

The test is crude and it works. Take three to five real buyer questions and put each of them to Doubao, DeepSeek, ERNIE Bot and ChatGPT. Check whether your numbers appear. Numbers without your brand name still count as half a win: the text was read, it simply has not been attributed to you yet.

In AI search optimization this is called a baseline measurement, and without one you cannot tell whether anything changed later. The method itself belongs to another piece.

Three questions buyers and owners actually ask

Can we publish a spec sheet that clients only receive on request?

Split it in two. General parameters — material range, size range, standard tolerance, certifications — carry no risk when published, because competitors already know them. What is genuinely sensitive is cost structure, customer-specific customization, and process recipes. Those stay off the website and out of the PDF. The test is not whether something can be public, but how much it costs you if a competitor reads it.

If we move PDF content to the web, will duplicate content hurt us?

The same content appearing in full at two URLs does split signals. The answer is division of labour rather than duplication: the page carries the summary and the reasoning, the PDF carries the full data and test methods, and the two do not repeat each other. If you genuinely need both, place a canonical on the PDF URL pointing to the HTML version.

We do not even have a proper product page. Where do we start?

Start with the three models that sell best. Give each one an HTML parameter page, with fields filled according to what buyers actually ask. Finish those three pages, then decide whether to extend. It is roughly a day of work, far cheaper than a new brochure, and AI can read it.

Closing

Back to the precision hardware factory. What they did was simple: they took the nine parameters most often asked about out of the 12-page PDF and put them in a table at the top of the product page; kept the PDF and moved its link below the table; then built two pure-HTML spec pages for their two flagship models.

When a buyer later asked the same tolerance question in DeepSeek, the answer contained can hold ±0.01 mm, requires wire EDM — with no company name attached.

But at least that tolerance no longer existed only inside a file you had to download.

Yihang began doing GEO in 2024; before that the business was cross-border B2B and Alibaba.com services. Product publishing on Alibaba.com has one characteristic worth noting: parameters are form fields. You fill them one at a time, and you cannot publish without them. It was inconvenient, but it forced sellers to turn parameters into structured text. A company website imposes no such constraint, so everything goes into a PDF — and what goes in is not a file. It is the part of you that AI cannot read.

Data and effect boundaries: statements here about PDF, Excel, Word and PowerPoint being indexable come from Google Search Central's "File types indexable by Google," including the phrase about requiring a specific parser to extract human-readable text, which is the documentation's own wording. That documentation describes indexability; it does not mean any particular file will parse successfully, nor that it will be cited in an AI answer. The 85% and 83% figures from AirOps are vendor-published observational statistics, not independently reviewed, and reflect a non-Chinese market snapshot; they should not be transferred directly to Doubao, DeepSeek or ERNIE Bot. The 12-page sheet, the nine parameters and the two flagship models are illustrative situations and describe no specific client. Every figure cited here is a mechanism description or an observational statistic; none of it predicts results for any business. The methods described make existing technical information readable and liftable. They promise no volume of inquiries, no ranking, and no commercial outcome.

References

1. Google Search Central: "File types indexable by Google" (PDF, Excel, Word, PowerPoint, PostScript and EPUB listed as encoded file types requiring a specific parser to extract human-readable text). https://developers.google.com/search/docs/crawling-indexing/indexable-file-types
2. AirOps: State of AI Search 2026 (85% of brand mentions from third-party pages; 83% of commercial citations from pages updated within a year; vendor-published, observational, non-Chinese market).
3. What Gets Cited, accepted at SIGIR 2026 (252,000 controlled experiments; topical relevance and context position dominate; peer reviewed).
4. arXiv:2410.22349 (quote completeness 20%–25%; not peer reviewed, laboratory conditions).